@vibe-agent-toolkit/agent-skills 0.1.42 → 0.2.0-rc.10

package/dist/skill-test/run-harness.js

@@ -3,7 +3,7 @@
  *
  * Implements the full harness sequence as a pure async function (no process.exit):
  *   lock → assert safe workdir + harness root → stage → resolve staged-subject
- *   eval path → bootstrap-check (scaffold template + exit 3 if absent)
+ *   eval path → bootstrap-check (scaffold template + exit 2 with `Reason: bootstrap` if absent)
  *   → preflight (return early exitCode 2 on failure) → ack enforcement
  *   → dry-run short-circuit → build per-eval work items → run the vat-owned
  *   executor→grader pipeline (bounded-parallel) → merge grader fragments →
@@ -11,27 +11,34 @@
  *   → release lock → return result
  */
 import { randomBytes } from 'node:crypto';
-import { existsSync, lstatSync, readFileSync, rmSync, writeFileSync } from 'node:fs';
+import { existsSync, lstatSync, readFileSync, readdirSync, renameSync, rmSync, writeFileSync } from 'node:fs';
+import { homedir } from 'node:os';
 import { basename, dirname } from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { getToolVersion, killAllActiveClaudeChildren, mkdirSyncReal, normalizedTmpdir, probeAuthStatus, resolveAssetReference, safeExecResult, safePath, toForwardSlash, } from '@vibe-agent-toolkit/utils';
+import { ExitCode } from '@vibe-agent-toolkit/schema';
+import { direntKind, isPathAbsentError, mkdirSyncReal, normalizedTmpdir, prefixMessageOnce, resolveAssetReference, safePath, toForwardSlashAnyPlatform, } from '@vibe-agent-toolkit/utils';
+import { getToolVersion, safeExecResult, } from '@vibe-agent-toolkit/utils/process';
+import { killAllActiveClaudeChildren, probeAuthStatus, } from '@vibe-agent-toolkit/utils/skill-test';
 import { resolveSkillSource } from '../skill-source/resolve-skill-source.js';
+import { BaselineArtifactSchema, computeBaselineDelta, formatBaselineDeltaLine, } from './baseline-delta.js';
+import { activeContaminationSignals, armExpectationSkew, detectBaselineContamination, scrubControlArmEnv, skillContentNeedles, summarizeBaselineIntegrity, } from './baseline-integrity.js';
 import { assembleChildEnv, assertKnownEnvTokens, computeEnvTokens, resolveInjectEnv } from './declared-env.js';
 import { runExecutorForEval } from './eval-executor.js';
-import { runGraderForEval } from './eval-grader.js';
-import { EvalInputError, parseEvalSuite, stageEvalWorkspaces } from './eval-inputs.js';
+import { GRADER_FRAGMENT_UNPARSEABLE, GraderFragmentUnparseableError, runGraderForEval } from './eval-grader.js';
+import { armDirSegment, EvalInputError, parseEvalSuite, stageEvalWorkspaces, } from './eval-inputs.js';
 import { lintEvalExpectations, lintToolExpectationExecutables } from './eval-lint.js';
 import { DEFAULT_EVALS_SUBPATH } from './eval-suite-isolation.js';
 import { writeEvalsTemplate } from './evals-template.js';
-import { BootstrapNeededError, DuplicateStagedSkillError, InternalHarnessError, SkillTestExitCode, } from './exit-codes.js';
+import { BootstrapNeededError, DuplicateStagedSkillError, InternalHarnessError, } from './failure-reason.js';
 import { mergeFragmentsToFriction, mergeFragmentsToGrading, mergeFragmentsToToolEval } from './fragment-merge.js';
 import { FrictionReportSchema } from './friction-schema.js';
 import { DEFAULT_CONCURRENCY, DEFAULT_GRADER_MODEL } from './grader-model.js';
+import { sanitizeGraderText } from './grader-text.js';
 import { reconcileGrading } from './grading-adapter.js';
 import { GradingReportSchema } from './grading-schema.js';
-import { assertSafeHarnessRoot, assertSafeWorkdir, prepareHarnessRoot, resolveHarnessRoot } from './harness-location.js';
+import { assertSafeHarnessRoot, assertSafeWorkdir, HarnessLocationError, prepareHarnessRoot, resolveHarnessRoot, } from './harness-location.js';
 import { acquireHarnessLock, installSignalCleanup } from './lock.js';
-import { runPipeline } from './pipeline.js';
+import { RateLimitSignal, runPipeline } from './pipeline.js';
 import { detectPluginLayout } from './plugin-layout.js';
 import { runPreflight } from './preflight.js';
 import { descriptorToSource, stageHarness } from './staging.js';
@@ -49,7 +56,7 @@ import { verifyVendoredManifest } from './vendor-manifest.js';
 const VENDORED_SKILL_CREATOR_DIR = safePath.resolve(dirname(fileURLToPath(import.meta.url)), '../../vendor/skill-creator');
 /**
  * Absolute path to the vendored skill-creator grader rubric each per-eval grader
- * spawn is told to judge against (issue #145). Lives inside
+ * spawn is told to judge against. Lives inside
  * {@link VENDORED_SKILL_CREATOR_DIR}, whose hash manifest is verified during
  * preflight — so the rubric the grader uses is the pinned, integrity-checked copy.
  */
@@ -82,28 +89,30 @@ export const SKILL_TEST_BUILTIN_CAPS = {
 };
 /**
  * Resolve the effective PER-EVAL wall-clock timeout (ms). Each executor and
- * grader spawn is an independent, bounded-parallel unit (issue #145), so the
+ * grader spawn is an independent, bounded-parallel unit, so the
  * budget is a flat per-spawn ceiling — NOT scaled by the suite size the way the
  * old single serial run's budget was. An explicit `--timeout` (seconds)
  * wins; otherwise the flat {@link DEFAULT_TIMEOUT_MS} default applies.
  */
-export function resolveTimeoutMs(opts) {
+function resolveTimeoutMs(opts) {
     return opts.timeout === undefined ? DEFAULT_TIMEOUT_MS : opts.timeout * 1000;
 }
 /**
  * Map an eval verdict to a process exit code. Default behavior (fail-closed): a
- * failing verdict escalates to EvalFailure (4) — distinct from the harness-broke
- * codes (1/2/3) so a CI consumer can `case $? in 0);; 4) tolerate;; *) hard fail;; esac`.
+ * failing verdict is `FINDINGS` — the harness ran to completion and what it
+ * examined did not pass, exactly as `vat skills validate` reports an
+ * error-severity finding — while a harness that could not run is `ERROR`, so a
+ * CI consumer can `case $? in 0);; 1) tolerate;; *) hard fail;; esac`.
  * When `tolerateEvalFailure` is set (interactive opt-out), a failing verdict is
- * downgraded to Ok (0) and the count lives only in the summary/grading.json.
+ * downgraded to `OK` and the count lives only in the summary/grading.json.
  */
 export function verdictExitCode(allPassed, tolerateEvalFailure) {
-    return !allPassed && !tolerateEvalFailure ? SkillTestExitCode.EvalFailure : SkillTestExitCode.Ok;
+    return !allPassed && !tolerateEvalFailure ? ExitCode.FINDINGS : ExitCode.OK;
 }
-export function resolveStallMs(opts) {
+function resolveStallMs(opts) {
     return opts.stall === undefined ? undefined : opts.stall * 1000;
 }
-export function resolveKnobs(opts) {
+function resolveKnobs(opts) {
     const result = {
         maxTurns: opts.maxTurns ?? DEFAULT_MAX_TURNS,
         maxBudgetUsd: opts.maxBudgetUsd ?? DEFAULT_MAX_BUDGET_USD,
@@ -123,13 +132,13 @@ export function resolveKnobs(opts) {
  * sources (npm/url/vendored/workspace) have no local source tree to walk, so they
  * are always staged flat (returns undefined). undefined → flat staging.
  */
-export function detectItemPluginLayout(source, repoRoot) {
+function detectItemPluginLayout(source, repoRoot) {
     if (!('path' in source))
         return undefined;
     const sourceDir = resolveAssetReference(source.path, repoRoot);
     return detectPluginLayout(sourceDir, existsSync) ?? undefined;
 }
-export function makeStageItem(name, source, repoRoot, role, optional) {
+function makeStageItem(name, source, repoRoot, role, optional) {
     const pluginLayout = detectItemPluginLayout(source, repoRoot);
     return {
         name,
@@ -143,14 +152,14 @@ export function makeStageItem(name, source, repoRoot, role, optional) {
  * Build the full set of items `stageHarness` will stage: the subject, every
  * REQUIRED `--with` companion, and every OPTIONAL `--with-optional` companion.
  * Both `with` and `optional` STAGE the named companion and make it invocable —
- * they differ only in required-vs-optional resolution (issue #153; a `--with`
- * name that named no positional skill used to be silently dropped, never staged,
- * no manifest trace). Fail-closed on a DUPLICATE staged name (subject / `with` /
+ * they differ only in required-vs-optional resolution (a `--with` name that
+ * names no positional skill must never be silently dropped — unstaged, with no
+ * manifest trace). Fail-closed on a DUPLICATE staged name (subject / `with` /
  * `optional` all share one namespace): the first repeat throws
  * {@link DuplicateStagedSkillError} rather than silently letting a later item
  * clobber an earlier one under the same staged slot.
  */
-export function buildStageItems(opts, repoRoot) {
+function buildStageItems(opts, repoRoot) {
     const items = [];
     const seen = new Set();
     const pushItem = (item) => {
@@ -173,7 +182,7 @@ export function buildStageItems(opts, repoRoot) {
     }
     return items;
 }
-export function buildResolveCtx(harnessRoot, repoRoot) {
+function buildResolveCtx(harnessRoot, repoRoot) {
     return {
         repoRoot,
         stagingRoot: safePath.joinUnderRoot(harnessRoot, 'staged'),
@@ -181,56 +190,111 @@ export function buildResolveCtx(harnessRoot, repoRoot) {
     };
 }
 /**
- * Map of flags to dummy values used when probing for flag support.
+ * Match a flag name as a whole token in `claude --help` output.
  *
- * Each value is chosen so value-validation (e.g., enum checks) accepts it
- * before `--help` short-circuits the session. The empty string `''` for
- * `--setting-sources` is intentional — an empty comma-separated list is valid.
+ * The boundary matters: a bare `includes('--plugin-dir')` also matches
+ * `--plugin-dirs` or `--plugin-dir-cache`, and vat's own flag names are prefixes
+ * of each other (`--max-turns` / `--max-turns-per-eval` would collide). Neither
+ * side of the token may be a letter or a dash.
  */
-const FLAG_DUMMY_VALUES = {
-    '--plugin-dir': '.',
-    '--setting-sources': '',
-    '--output-format': 'stream-json',
-    '--permission-mode': 'bypassPermissions',
-    '--max-turns': '1',
-    '--max-budget-usd': '1',
-};
+function helpTextDeclaresFlag(helpText, flag) {
+    // A scan rather than a built regex: the flag is data (it comes from a list a
+    // future edit will extend), and building a pattern from data is both a lint
+    // error here and the kind of thing that silently changes meaning when someone
+    // adds a flag containing a metacharacter.
+    const isTokenChar = (ch) => ch !== undefined && (ch === '-' || ch === '_' || /[a-z0-9]/i.test(ch));
+    for (let from = helpText.indexOf(flag); from !== -1; from = helpText.indexOf(flag, from + 1)) {
+        if (!isTokenChar(helpText[from - 1]) && !isTokenChar(helpText[from + flag.length]))
+            return true;
+    }
+    return false;
+}
 /**
- * Return a sensible dummy value for a given CLI flag so that value-validation
- * doesn't reject the argument before `--help` can short-circuit the session.
+ * A flag name no `claude` will ever accept. It is probed alongside the real ones
+ * as a NEGATIVE CONTROL: if the probe reports this as supported, the probe is
+ * broken and its answer about every other flag is worthless.
  */
-export function flagDummyValueFor(flag) {
-    return FLAG_DUMMY_VALUES[flag] ?? '1';
-}
+const FLAG_PROBE_SENTINEL = '--vat-probe-flag-that-cannot-exist';
 /**
- * Build a token-free flag-parse probe.
+ * Build a token-free flag-support probe from `claude --help`.
+ *
+ * ⚠️ This REPLACES an exit-code probe (`claude <flag> <dummy> --help`, exit 0 ⇒
+ * supported) that could not discriminate at all. Verified against claude 2.x:
+ * `claude --no-such-flag-xyz 1 --help` exits **0**, because `--help` short-circuits
+ * before argument validation. Every one of preflight's `flag <name>` checks
+ * therefore reported "supported" unconditionally — including for flags that do not
+ * exist — so the gate that is supposed to stop vat spawning with an unsupported
+ * flag was decorative. (The same command without `--help` exits 1, but running it
+ * would start a real session and bill tokens, which is why `--help` was there.)
  *
- * For each flag, runs `claude <flag> <dummyValue> --help` via safeExecResult.
- * Exit 0 means claude's parser accepts the flag; `--help` short-circuits before
- * any session or tokens are consumed.
+ * Help-text matching does discriminate, and it is still token-free: one `--help`
+ * invocation for the whole run instead of one spawn per flag.
  *
- * The dummy value is always included in the args, even when it is an empty
- * string — some flags (e.g., `--setting-sources`) require a value argument,
- * so omitting it causes exit 1 before `--help` can be reached.
+ * The probe carries its own negative control ({@link FLAG_PROBE_SENTINEL}). If a
+ * future `claude` prints something that matches the sentinel — or `--help` fails
+ * outright — the probe reports EVERY flag unsupported rather than every flag
+ * supported. A probe that cannot tell must fail closed and be seen; the defect it
+ * replaces failed open and was invisible.
  *
- * The probe is a closure — it is NOT spawned at module scope, only at call time
- * inside runPreflight.
+ * `--help` runs on the FIRST flag query, not when the probe is built. Building
+ * it is part of assembling preflight's input, and a preflight that never asks
+ * (a stubbed one, or one that fails an earlier check) must not have already
+ * paid for a subprocess: with the spawn at build time every harness run cost
+ * one real `claude --help` — 75 ms where the CLI is installed, an instant
+ * ENOENT where it is not — which read as a 10× platform split in a test that
+ * had injected a fake spawn and stubbed preflight precisely so no process
+ * would start.
  */
-function buildFlagParseProbe() {
+function buildFlagParseProbe(runHelp = defaultClaudeHelp) {
+    let helpText;
     return (flag) => {
-        const dummy = flagDummyValueFor(flag);
-        const result = safeExecResult('claude', [flag, dummy, '--help'], {
-            stdio: 'pipe',
-            encoding: 'utf8',
-            timeout: 15_000,
-        });
-        return result.success;
+        if (helpText === undefined)
+            helpText = runHelp();
+        // No help output, or a sentinel "match" ⇒ the probe is not trustworthy.
+        const usable = helpText !== null && !helpTextDeclaresFlag(helpText, FLAG_PROBE_SENTINEL);
+        return usable && helpTextDeclaresFlag(helpText ?? '', flag);
     };
 }
-export function buildPreflightInput(evalsPath, pluginDirs, opts, knobs) {
+/** Run `claude --help` once, returning its combined output (null if unreachable). */
+function defaultClaudeHelp() {
+    const result = safeExecResult('claude', ['--help'], {
+        stdio: 'pipe',
+        encoding: 'utf8',
+        timeout: 15_000,
+    });
+    if (!result.success)
+        return null;
+    // `stdout`/`stderr` are typed `Buffer | string`; `encoding: 'utf8'` makes them
+    // strings, and `String()` is correct either way. Both streams are read because
+    // a CLI is free to print its usage to either.
+    return `${String(result.stdout)}\n${String(result.stderr)}`;
+}
+/**
+ * ⚠️ KNOWN OPEN — `costEstimate` here is DEAD, in three independent ways, and none
+ * of them is fixable from inside this function:
+ *
+ * 1. `evalCount: 1` is a placeholder. The suite is not parsed until Step 5.5, which
+ *    runs AFTER preflight, so the real count does not exist yet. Making it real means
+ *    either moving preflight after the suite parse (it currently gates that parse) or
+ *    adding a second reader of `evals.json` — and two readers of the answer key is
+ *    precisely the drift this module refuses elsewhere.
+ * 2. `renderPreflightSummary` returns only FAILING checks, and it is called only
+ *    inside the `if (!preflightResult.passed)` branch — so on a passing run the
+ *    estimate never reaches a terminal at all.
+ * 3. Preflight runs after the `--dry-run` short-circuit, so a dry run never gets here.
+ *
+ * The consequence is that {@link buildDryRunSummary}'s spawn-count line is the ONLY
+ * pre-spend number an operator ever sees. Fix the count before making the check
+ * reachable: a reachable estimate that says "1 evals × 2 config" is worse than a
+ * silent one, because it is quotable.
+ */
+function buildPreflightInput(evalsPath, pluginDirs, opts, knobs) {
     const costEstimate = {
         evalCount: 1,
-        configurations: 1,
+        // `configurations` IS the A/B dimension: `--baseline` runs every eval under two
+        // configurations (skill declared, skill withheld). Correct even while
+        // `evalCount` is a placeholder, and it is the half that does not need the suite.
+        configurations: opts.baseline === true ? 2 : 1,
         runsPerQuery: 1,
         maxBudgetUsd: knobs.maxBudgetUsd,
     };
@@ -256,7 +320,7 @@ export function buildPreflightInput(evalsPath, pluginDirs, opts, knobs) {
         preflightOpts.requireAuth = opts.requireAuth;
     return preflightOpts;
 }
-export function renderPreflightSummary(checks) {
+function renderPreflightSummary(checks) {
     const failed = checks.filter(c => !c.passed);
     return failed.length > 0
         ? failed.map(c => `  [FAIL] ${c.name}: ${c.message}`).join('\n')
@@ -272,42 +336,79 @@ export function renderPreflightSummary(checks) {
 export function isAcknowledged(opts) {
     return opts.dryRun === true || opts.acknowledgedRunsSkillCode === true;
 }
+/**
+ * Most friction lines vat will put on an operator's terminal in one run. The
+ * full set is always on disk; a grader that emits thousands of items must not
+ * be able to scroll the run's real output away.
+ */
+const MAX_FRICTION_LINES = 50;
 /**
  * Format a friction report for human consumption — one line per entry as
  * `[<severity>] <category>: <message>`. Pure; returns the empty string for no
  * entries so the caller can skip emitting anything.
+ *
+ * `severity` and `category` are closed enums vat owns, so they are safe to
+ * interpolate. `message` is GRADER text and is sanitized here even though
+ * {@link parseEvalFragment} already sanitized it on the way in — this is a
+ * second, independent boundary, not a redundant one: {@link emitFrictionReport}
+ * re-reads `friction.json` from the harness results dir, which same-uid skill
+ * code can reach and rewrite after vat wrote it.
  */
 export function formatFrictionReport(items) {
-    return items.map(i => `[${i.severity}] ${i.category}: ${i.message}`).join('\n');
+    const shown = items.slice(0, MAX_FRICTION_LINES);
+    const lines = shown.map(i => `[${i.severity}] ${i.category}: ${sanitizeGraderText(i.message)}`);
+    if (items.length > shown.length) {
+        lines.push(`... and ${items.length - shown.length} more (full list in friction.json)`);
+    }
+    return lines.join('\n');
 }
 /**
  * Read the run's friction.json (if present + valid) and echo a concise report to
  * STDERR so users don't miss packaging-fidelity friction VAT merged from the grader
- * fragments (it is otherwise only written to disk). Best-effort: a missing, unparseable, or
- * empty report emits nothing. Never touches stdout (which stays machine-readable).
- * Accepts `undefined` (a no-op) so the harness `finally` can call it unconditionally
- * even when a throw preempted assignment of the friction path.
+ * fragments (it is otherwise only written to disk). Best-effort: a missing or empty
+ * report emits nothing. A report that is there but cannot be read or does not parse
+ * emits a one-line warning instead — this runs from the harness `finally`, so it
+ * must not throw, but vat is the sole writer of friction.json and a file it wrote
+ * that it cannot read back is worth a line. Never touches stdout (which stays
+ * machine-readable). Accepts `undefined` (a no-op) so the harness `finally` can
+ * call it unconditionally even when a throw preempted assignment of the path.
  */
 function emitFrictionReport(frictionPath) {
     if (frictionPath === undefined)
         return;
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
     if (!existsSync(frictionPath))
         return;
     let raw;
     try {
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
         raw = JSON.parse(readFileSync(frictionPath, 'utf-8'));
     }
-    catch {
+    catch (err) {
+        process.stderr.write(`warning: could not read the friction report at ${frictionPath}: ` +
+            `${err instanceof Error ? err.message : String(err)}\n`);
         return;
     }
     const parsed = FrictionReportSchema.safeParse(raw);
-    if (!parsed.success || parsed.data.items.length === 0)
+    if (!parsed.success) {
+        process.stderr.write(`warning: the friction report at ${frictionPath} does not match its schema: ${parsed.error.message}\n`);
+        return;
+    }
+    if (parsed.data.items.length === 0)
         return;
     process.stderr.write(`\nPackaging friction (${parsed.data.items.length}):\n`);
     process.stderr.write(formatFrictionReport(parsed.data.items) + '\n');
 }
+/**
+ * Write the `--baseline` delta line and ⚠️ banner to stderr, if this run produced
+ * one. A no-op on a non-baseline run.
+ *
+ * A function rather than an `if` at the call site purely to keep the orchestrator
+ * inside its cognitive-complexity budget; the ORDER it is called in is the part
+ * that matters, and that is documented where it is called.
+ */
+function emitBaselineReport(report) {
+    if (report !== undefined)
+        process.stderr.write(report);
+}
 /**
  * Name any optional companions (`--with-optional`/config `optional:`) that were
  * skipped because their source could not be resolved — never silently, so the
@@ -317,7 +418,8 @@ function emitFrictionReport(frictionPath) {
 function emitSkippedOptionalWarning(skippedOptional) {
     if (skippedOptional.length === 0)
         return;
-    process.stderr.write(`warning: optional companion skill(s) not staged (source unresolvable): ${skippedOptional.join(', ')}\n`);
+    const named = skippedOptional.map((item) => `${item.name} (${item.reason})`).join(', ');
+    process.stderr.write(`warning: optional companion skill(s) not staged: ${named}\n`);
 }
 /**
  * Resolve the PERSISTENT location where a bootstrap scaffold should be written
@@ -326,8 +428,8 @@ function emitSkippedOptionalWarning(skippedOptional) {
  * that real source dir; otherwise we anchor under the repo root by skill name.
  */
 /** The subject skill's display name (trailing segment of the subject arg). */
-export function subjectSkillName(opts) {
-    return basename(toForwardSlash(opts.subject)) || opts.subject;
+function subjectSkillName(opts) {
+    return basename(toForwardSlashAnyPlatform(opts.subject)) || opts.subject;
 }
 /**
  * Where this run READS its eval suite from, anchored at the subject's authored
@@ -351,7 +453,7 @@ export function subjectSkillName(opts) {
  * into `<skillDir>/<path>`, which does not exist — so the run did not fail, it
  * bootstrapped a starter template at the bogus location and graded that.
  */
-export function resolveScaffoldEvalsPath(opts, repoRoot, evalsRef) {
+function resolveScaffoldEvalsPath(opts, repoRoot, evalsRef) {
     const resolveFrom = (baseDir) => evalsRef === undefined
         ? safePath.join(baseDir, DEFAULT_EVALS_SUBPATH)
         : resolveAssetReference(evalsRef, baseDir);
@@ -370,7 +472,7 @@ export function resolveScaffoldEvalsPath(opts, repoRoot, evalsRef) {
     return resolveFrom(sourceDir);
 }
 /**
- * Bootstrap (exit 3) — called when the eval suite is absent everywhere. A real run
+ * Bootstrap (exit 2, `Reason: bootstrap`) — called when the eval suite is absent everywhere. A real run
  * writes a persistent template at the source scaffold location and reports it; a dry
  * run must never touch the filesystem, so it reports where a real run *would*
  * scaffold and writes nothing. Both surface the same exit-3 BootstrapNeededError.
@@ -391,7 +493,7 @@ function bootstrapEvalSuite(opts, repoRoot, evalsRef) {
  * is performing. Created 0700 and removed in the run's cleanup, exactly like
  * {@link resolveGraderOutDir}. Pure (derives a path only).
  */
-export function resolveEvalSuiteHoldDir(dirToken) {
+function resolveEvalSuiteHoldDir(dirToken) {
     return safePath.join(normalizedTmpdir(), `vat-skill-evals-${dirToken}`);
 }
 /**
@@ -404,38 +506,78 @@ export function resolveEvalSuiteHoldDir(dirToken) {
  * 2. **The vat-only hold dir**, when staging harvested a suite that exists nowhere
  *    else (a fetched artifact). Its layout mirrors an authored evals dir, so
  *    `fixtures/` resolve relative to it unchanged.
- * 3. Neither → `undefined`, and the caller bootstraps a template (exit 3).
+ * 3. Neither → `undefined`, and the caller bootstraps a template (exit 2, `Reason: bootstrap`).
  *
  * Returns the suite's absolute path; its `dirname` is the base for each eval's
  * declared input `files`.
  */
-export function resolveEvalSuitePath(input) {
+function resolveEvalSuitePath(input) {
     const authored = resolveScaffoldEvalsPath(input.opts, input.repoRoot, input.evalsRef);
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- authored source path resolved from opts
     if (existsSync(authored))
         return authored;
     if (!input.subjectEvalSuiteHeld)
         return undefined;
     return safePath.join(input.holdDir, basename(input.evalsSubpath));
 }
+/**
+ * The per-run executor workspace root: `<tmp>/vat-skill-test-ws-<dirToken>/`.
+ *
+ * Deliberately OUTSIDE `harnessRoot`, alongside {@link resolveGraderOutDir} and
+ * {@link resolveEvalSuiteHoldDir}. The executor's cwd lives under here, and the
+ * harness root is where vat stages runnable copies of the skill (`staged/` and
+ * the assembled plugin dir). While the two were parent and child, the skill-absent
+ * arm of a `--baseline` run sat one `ls ..` away from the treatment it was
+ * supposed to be denied — the control could reach vat's own copy without ever
+ * leaving its working directory.
+ *
+ * Named by an unpredictable token, like the other vat-only dirs, and removed by
+ * the same cleanup. Pure (derives a path only).
+ */
+function resolveWorkspacesRoot(dirToken) {
+    return safePath.join(normalizedTmpdir(), `vat-skill-test-ws-${dirToken}`);
+}
 /** Parse the staged eval suite and materialize each eval's input `files` into
- * `<harnessRoot>/workspaces/<id>/`. Returns the workspaces root, the parsed
- * {@link EvalSuite} (so the eval loop has the entries without re-reading), and the
- * declared eval count (derived from the suite). The dir is wiped first so a reused
- * harness root cannot leak a prior run's inputs. Throws {@link EvalInputError}
- * (mapped by the caller to exit 2) on a bad suite or a missing input file. */
-export function stageWorkspacesForRun(evalsPath, harnessRoot) {
-    const workspacesRoot = safePath.joinUnderRoot(harnessRoot, 'workspaces');
+ * `<workspacesRoot>/<arm>/<id>/`, once per arm this run will execute. Returns the
+ * workspaces root, the parsed {@link EvalSuite} (so the eval loop has the entries
+ * without re-reading), and the declared eval count (derived from the suite). The
+ * dir is wiped first so a reused root cannot leak a prior run's inputs. Throws
+ * {@link EvalInputError} (mapped by the caller to exit 2) on a bad suite or a
+ * missing input file. */
+/**
+ * Mint this run's opaque per-arm workspace directory segments.
+ *
+ * Independent random tokens, NOT the arm names and not derived from one another:
+ * the segment ends up in the executor's cwd, which the prompt states verbatim, so
+ * a name like `without` unblinds the control arm outright — and a name like `with`
+ * hands it a guessable path to the treatment arm's live output one directory over.
+ * 8 bytes is ample for "not guessable within a run that lasts minutes"; these are
+ * blinding/anti-enumeration tokens, not secrets protecting data at rest.
+ */
+function mintArmWorkspaceDirs(baseline) {
+    const token = () => randomBytes(8).toString('hex');
+    return baseline ? { with: token(), without: token() } : { with: token() };
+}
+function stageWorkspacesForRun(evalsPath, workspacesRoot, armDirs) {
     rmSync(workspacesRoot, { recursive: true, force: true });
     // 0700, matching the harness root. These dirs hold each eval's declared input
     // files, which with an out-of-tree suite may be data that was never in the repo
     // at all. Inheriting the umask (0755) left them readable by any local user the
     // moment `--out` relocated the harness root out from under its 0700 parent.
     mkdirSyncReal(workspacesRoot, { recursive: true, mode: 0o700 });
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- evalsPath is our staged-subject path
+    // The same shared-tmp hardening the harness root gets. `mkdirSync(recursive)`
+    // on an existing path neither throws nor chmods, so without this an attacker
+    // winning the race between the rmSync and the mkdir owns the executor's working
+    // directory and nothing ever re-checks. Cheap: the leaf sits directly under the
+    // trusted tmp boundary, so the ancestry walk degrades to one lstat + stat.
+    assertSafeHarnessRoot(workspacesRoot, process.getuid?.() ?? -1);
     const suite = parseEvalSuite(readFileSync(evalsPath, 'utf-8'));
     return {
-        workspacesRoot: stageEvalWorkspaces({ suite, evalsDir: dirname(evalsPath), workspacesRoot }),
+        workspacesRoot: stageEvalWorkspaces({
+            suite,
+            evalsDir: dirname(evalsPath),
+            workspacesRoot,
+            armDirs,
+        }),
         declaredEvalCount: suite.evals.length,
         suite,
     };
@@ -452,15 +594,16 @@ function buildModelFlag(model) {
  * Keeping the try/catch in a private helper avoids inflating the orchestrator's
  * cognitive complexity.
  */
-function attemptStageWorkspaces(evalsPath, harnessRoot) {
+function attemptStageWorkspaces(evalsPath, harnessRoot, workspacesRoot, armDirs) {
     try {
-        return stageWorkspacesForRun(evalsPath, harnessRoot);
+        return stageWorkspacesForRun(evalsPath, workspacesRoot, armDirs);
     }
     catch (e) {
         if (e instanceof EvalInputError) {
             return {
                 harnessPath: harnessRoot,
-                exitCode: SkillTestExitCode.Preflight,
+                exitCode: ExitCode.ERROR,
+                reason: 'preflight',
                 summary: `Eval input error:\n  ${e.message}`,
             };
         }
@@ -470,8 +613,8 @@ function attemptStageWorkspaces(evalsPath, harnessRoot) {
 /**
  * Advisory (never-fatal) lint pass over the parsed suite: nudges authors away
  * from presence-only expectations ("mentions/includes …") that a hallucinated
- * or wrong-for-the-right-reason answer could still satisfy (issue #145
- * follow-up). Purely additive stderr noise, emitted before any spawn — never
+ * or wrong-for-the-right-reason answer could still satisfy.
+ * Purely additive stderr noise, emitted before any spawn — never
  * affects exitCode. Extracted so the Step 5.5 call site stays a single line
  * (keeps the orchestrator's cognitive complexity budget).
  */
@@ -575,6 +718,38 @@ export function buildStaleDistWarningLines(roleLabel) {
  * Always includes the assembled spawn command, the staged-manifest entry count +
  * fingerprint, and the provenance.json path so a stale tree is visible at a glance.
  */
+/** `s` unless the count is exactly one — the plural the two spawn-plan lines share. */
+function plural(count) {
+    return count === 1 ? '' : 's';
+}
+/**
+ * The two lines that say WHAT A REAL RUN WOULD COST, in PAIRS — not evals.
+ *
+ * `--baseline` runs every eval on both arms and each arm is a full executor→grader
+ * pair, so the spawn count is `evals × arms` and the `claude` session count is twice
+ * that. Reporting the eval count as the pair count understated a baseline run by
+ * exactly 2x: a 3-eval suite previewed as "3 evals as executor→grader spawn pairs"
+ * and then launched 6 pairs / 12 sessions.
+ *
+ * The arm count is spelled out rather than folded silently into the number, because
+ * "6 pairs" for a 3-eval suite reads as a bug until you are told why. And the budget
+ * flag is labelled PER SPAWN with the ceiling multiplied out, because the name
+ * `--max-budget-usd` reads as a run total — under `--baseline` that misreading
+ * under-provisions by a factor of four times the eval count.
+ */
+function buildSpawnPlanLines(input) {
+    const pairs = input.evalCount * (input.baseline ? 2 : 1);
+    const sessions = pairs * 2;
+    const armNote = input.baseline
+        ? ` (${input.evalCount} eval${plural(input.evalCount)} × 2 arms — --baseline runs each eval with AND without the skill)`
+        : '';
+    return [
+        `[dry-run] Would run ${pairs} executor→grader spawn pair${plural(pairs)}${armNote} ` +
+            `at concurrency ${input.concurrency} — ${sessions} claude session${plural(sessions)} in total.`,
+        `[dry-run] --max-budget-usd is PER SPAWN ($${input.maxBudgetUsd}), not per run: worst case ≈ ` +
+            `$${(sessions * input.maxBudgetUsd).toFixed(2)} across those ${sessions} sessions.`,
+    ];
+}
 export function buildDryRunSummary(input) {
     const lines = [];
     if (input.wouldBuild) {
@@ -604,41 +779,71 @@ export function buildDryRunSummary(input) {
     // it is labelled rather than left to be taken at face value.
     const stagedFromUnbuiltTree = input.wouldBuild && input.dryRunStagedExistingDist !== undefined;
     const provisional = stagedFromUnbuiltTree ? ' (PROVISIONAL — not rebuilt)' : '';
-    lines.push(`[dry-run] Would run ${input.evalCount} eval${input.evalCount === 1 ? '' : 's'} as executor→grader spawn ` +
-        `pair${input.evalCount === 1 ? '' : 's'} at concurrency ${input.concurrency}.`, `[dry-run] Executor ${input.modelFlag}; grader model ${input.graderModel} (prompt via stdin).`, `[dry-run] Staged manifest: ${count} entr${count === 1 ? 'y' : 'ies'} | ` +
-        `fingerprint: ${input.provenanceFingerprint}${provisional}`, `[dry-run] Provenance: ${input.provenancePath}`);
+    lines.push(...buildSpawnPlanLines(input), `[dry-run] Executor ${input.modelFlag}; grader model ${input.graderModel} (prompt via stdin).`, `[dry-run] Staged manifest: ${count} entr${count === 1 ? 'y' : 'ies'} | ` +
+        `fingerprint: ${input.provenanceFingerprint}${provisional}`, `[dry-run] Provenance would be written to: ${input.provenancePath}`);
     return lines.join('\n');
 }
 /**
- * Remove the harness directory after a run so staged untrusted skill bytes and
- * prompts do not accumulate in OS tmp. No-op when the user asked to keep it, when
- * the dir is a user-supplied location (`--out`/`--workdir`), or when it is already
- * gone. Idempotent and never throws — it runs from a `finally`, so it must not
- * mask the run's real outcome.
+ * The one child of the harness root that cleanup NEVER removes. Cleanup exists to
+ * evict staged untrusted skill bytes and prompts from OS tmp; `results/` is the
+ * opposite — it is the run's product, written solely by vat.
+ */
+const RETAINED_RESULTS_DIRNAME = 'results';
+/**
+ * Remove the harness directory's staged contents after a run so untrusted skill
+ * bytes and prompts do not accumulate in OS tmp. No-op when the user asked to keep
+ * it, when the dir is a user-supplied location (`--out`/`--workdir`), or when it is
+ * already gone. Idempotent and never throws — it runs from a `finally`, so it must
+ * not mask the run's real outcome.
+ *
+ * `results/` SURVIVES. It holds `grading.json`, `friction.json`, `tool-eval.json`
+ * and — for a `--baseline` run — `baseline.json`, i.e. every artifact the command's
+ * own help text tells the operator to go read. Removing the whole root on the
+ * DEFAULT invocation (no `--out`, no `--workdir`, no `--keep`) deleted those files
+ * inside this function's own `finally`, before `run.ts` had printed a single line:
+ * the operator was handed a `Harness:` path to a directory that no longer existed.
+ * That is the invocation `vat-skill-testing.md`'s copy-paste example uses.
+ *
+ * The retention is deliberately UNCONDITIONAL rather than `--baseline`-only: every
+ * run's verdict detail lives in the same directory, and a rule that depends on a
+ * flag is a rule that will drift away from the flag. The cost is bounded — the
+ * harness root is a deterministic function of the subject set, so a re-run reuses
+ * (and {@link wipeStaleArtifacts} clears) the same `results/`, rather than
+ * accumulating one per run.
  *
  * SAFETY: re-asserts the root is not a symlink immediately before removal (via
  * `lstat`, which does NOT follow the link). A root swapped to a symlink between
  * the run and cleanup is left in place rather than followed — `rmSync(recursive)`
  * could otherwise delete the symlink's target outside tmp.
  */
-export function cleanupHarness(harnessRoot, opts) {
+function cleanupHarness(harnessRoot, opts) {
     if (opts.keep || !opts.created)
         return;
     // Best-effort: cleanup runs from a `finally`, so a TOCTOU race (the dir is
     // reaped between checks) or a permission error must never throw out and mask
-    // the run's real outcome. Worst case is a leftover 0700 tmp dir, not a failure.
-    try {
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived harness root
+    // the run's real outcome. Worst case is a leftover 0700 tmp dir, not a failure —
+    // but a leftover the operator was never told about is staged untrusted bytes
+    // sitting in tmp with nothing to connect them to this run, so the failure is
+    // REPORTED on stderr ({@link swallowCleanupFailure}), not swallowed silently.
+    swallowCleanupFailure(() => {
         if (!existsSync(harnessRoot))
             return;
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived harness root
         if (lstatSync(harnessRoot).isSymbolicLink())
             return;
-        rmSync(harnessRoot, { recursive: true, force: true });
-    }
-    catch {
-        // Swallow: a failed cleanup is not a run failure.
-    }
+        const entries = readdirSync(harnessRoot);
+        // Nothing to retain — a run that ended before Step 7 (preflight refusal, lock
+        // failure, a throw during staging) has no results/, so leaving an empty 0700
+        // dir behind in tmp would be pure litter. Remove the root outright.
+        if (!entries.includes(RETAINED_RESULTS_DIRNAME)) {
+            rmSync(harnessRoot, { recursive: true, force: true });
+            return;
+        }
+        for (const entry of entries) {
+            if (entry === RETAINED_RESULTS_DIRNAME)
+                continue;
+            rmSync(safePath.joinUnderRoot(harnessRoot, entry), { recursive: true, force: true });
+        }
+    });
 }
 /**
  * Build the per-eval work items for a set of evals (one tier's worth, or a whole
@@ -648,7 +853,7 @@ export function cleanupHarness(harnessRoot, opts) {
  * (see {@link partitionFragmentsByArm}) and never drives tier gating. Pure +
  * unit-testable.
  */
-export function buildEvalWorkItems(evals, baseline) {
+function buildEvalWorkItems(evals, baseline) {
     const items = [];
     for (const entry of evals) {
         items.push({ entry, arm: 'with' });
@@ -663,7 +868,7 @@ export function buildEvalWorkItems(evals, baseline) {
  * fragment with no `arm` (or `arm: 'with'`) is a WITH-arm fragment. Pure +
  * unit-testable.
  */
-export function partitionFragmentsByArm(fragments) {
+function partitionFragmentsByArm(fragments) {
     const withArm = [];
     const withoutArm = [];
     for (const fragment of fragments) {
@@ -694,18 +899,47 @@ export function partitionFragmentsByArm(fragments) {
  * container; that is tracked as a follow-up (see CHANGELOG "Security" notes) and
  * is the only thing that closes the residual race outright.
  */
-export function resolveGraderOutDir(dirToken) {
+function resolveGraderOutDir(dirToken) {
     return safePath.join(normalizedTmpdir(), `vat-skill-grade-${dirToken}`);
 }
 /**
- * The executor working directory for one eval: its staged input workspace
- * `<workspacesRoot>/<id>` when the eval declares input `files`, else undefined
- * (the executor then defaults to the staged subject dir). Pure + unit-testable.
+ * The executor working directory for one eval ON ONE ARM: its staged input
+ * workspace `<workspacesRoot>/<arm>/<id>`, ALWAYS — empty when the eval declares
+ * no input `files` (see {@link stageEvalWorkspaces}).
+ *
+ * This used to return undefined for a file-less eval, which made the executor
+ * fall back to running inside the staged subject dir. That fallback was the
+ * quiet half of the `--baseline` control defect: the skill-absent arm's cwd
+ * was the skill.
+ *
+ * The per-arm segment closes the OTHER half, found a round later: without it both
+ * arms of one eval shared a single writable directory AND ran concurrently, so
+ * the control arm could read the treatment arm's output files and answer from
+ * them. Keyed on the eval id alone this is a pure function of the eval; the arm
+ * is what makes the two runs independent.
+ *
+ * The segment is an OPAQUE per-run token, never the arm's name — this path is
+ * quoted verbatim to the executor as its working directory, so `…/without/e1`
+ * told the control arm it was the control, and `../with/e1` told it where to look.
+ * See {@link mintArmWorkspaceDirs}. Pure + unit-testable.
  */
-export function resolvePerEvalWorkspaceDir(entry, workspacesRoot) {
-    if (entry.files === undefined || entry.files.length === 0)
-        return undefined;
-    return safePath.joinUnderRoot(workspacesRoot, String(entry.id));
+function resolvePerEvalWorkspaceDir(entry, workspacesRoot, arm, armDirs) {
+    return safePath.joinUnderRoot(armWorkspaceRoot(workspacesRoot, armDirs, arm), String(entry.id));
+}
+/**
+ * `<workspacesRoot>/<that arm's opaque segment>` — ONE arm's staged workspace root,
+ * derived in exactly one place.
+ *
+ * The consolidation is the point, not the line count. Every consumer that wants an
+ * arm root wants this join, and a consumer that takes the result as a bare `string`
+ * has no way to tell an arm root from the workspaces ROOT one level above it:
+ * handing {@link resolveSkillContentNeedles} `workspacesRoot` instead of
+ * `<workspacesRoot>/<withSegment>` type-checked, ran, and left all 2270 unit tests
+ * green while every fixture exclusion silently missed. Callers now pass the two
+ * inputs and let this derive the path.
+ */
+function armWorkspaceRoot(workspacesRoot, armDirs, arm) {
+    return safePath.joinUnderRoot(workspacesRoot, armDirSegment(armDirs, arm));
 }
 /**
  * Best-effort removal of a vat-only tmp dir that lives OUTSIDE the harness root —
@@ -713,57 +947,160 @@ export function resolvePerEvalWorkspaceDir(entry, workspacesRoot) {
  * cleanup, where masking the run's real outcome would be worse than a leftover 0700
  * tmp dir) and refuses to follow a symlinked root.
  */
-export function removeVatOnlyDir(dir) {
+/**
+ * Run one cleanup step, swallowing whatever it throws.
+ *
+ * The discipline every other step in the harness `finally` already follows, stated
+ * once: a cleanup failure is not a run failure, and letting one out REPLACES an
+ * already-good result — verdict computed, artifacts written, summary composed — with
+ * an error and exit 1.
+ *
+ * ⚠️ A CALL-SITE GUARD IS THE WEAK FORM, and this helper is a last resort rather
+ * than the pattern to copy. It is invisible to the compiler and to the test suite:
+ * reverting either of its two original call sites to a bare call left the unit suite
+ * AND the integration suite green, so nothing but review stood between the fix and
+ * its own silent removal. `lock.release()` was therefore made non-throwing at its own
+ * definition instead (see {@link acquireHarnessLock}), which is a contract the type
+ * carries and no edit here can undo. That leaves ONE caller —
+ * `killAllActiveClaudeChildren`, which lives in `@vibe-agent-toolkit/utils` and is
+ * only contractually non-throwing as long as that package says so.
+ *
+ * The failure is reported on stderr rather than silently dropped: a lockfile that
+ * could not be removed will break the NEXT run with a "busy" error, and an operator
+ * who saw nothing here has no way to connect the two.
+ */
+function swallowCleanupFailure(step) {
+    try {
+        step();
+    }
+    catch (err) {
+        process.stderr.write(`warning: harness cleanup step failed (the run's result stands): ${err instanceof Error ? err.message : String(err)}\n`);
+    }
+}
+/**
+ * Remove one vat-only temp dir. Same discipline as {@link cleanupHarness}: a failed
+ * removal is not a run failure, and it is reported rather than swallowed.
+ */
+function removeVatOnlyDir(dir) {
     if (dir === undefined)
         return;
-    try {
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived vat-only tmp dir
+    swallowCleanupFailure(() => {
         if (!existsSync(dir))
             return;
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived vat-only tmp dir
         if (lstatSync(dir).isSymbolicLink())
             return;
         rmSync(dir, { recursive: true, force: true });
-    }
-    catch {
-        // Swallow: a failed cleanup is not a run failure.
-    }
+    });
 }
 /**
  * Run ONE work item: the blind executor spawn, then the grader spawn over its
- * captured transcript (issue #145). Returns the grader fragment tagged with the
+ * captured transcript. Returns the grader fragment tagged with the
  * item's arm. The WITHOUT arm runs the executor with `pluginDirs: []` (skill
  * absent). Grader fragments are written under a PER-ARM subdir of the vat-only
  * grader dir so a WITH and WITHOUT run of the same eval id cannot collide.
  *
- * Throws propagate: an executor/grader {@link InternalHarnessError} (timeout,
- * stall, spawn error, grader failure/missing-fragment) fails the whole run
- * (exit 1); a RateLimitSignal is retried by the pipeline. An executor CLEAN
- * failure is NOT thrown — its transcript flows into the grader, whose failing
- * fragment surfaces as an eval failure (exit 4 via the verdict), never exit 1.
+ * Throws, and the caller ({@link runEvalWorker}) decides what a throw costs — see
+ * its docblock for why that depends on the arm. An executor CLEAN failure is NOT
+ * thrown on either arm: its transcript flows into the grader, whose failing
+ * fragment surfaces as an eval failure (exit FINDINGS via the verdict), never ERROR.
  */
-async function runEvalWorker(item, ctx) {
+/**
+ * The executor env for ONE eval.
+ *
+ * `workspaceDir` is threaded to the token resolver ONLY when the eval actually
+ * declares input `files`, even though every eval now HAS a workspace. The two are
+ * different questions: the workspace is where the executor runs, while
+ * `${fixturesDir}` promises a staged `fixtures/` directory that EXISTS — and
+ * `stageEvalWorkspaces` creates that only for evals declaring files. Passing it
+ * unconditionally silently downgraded `UnresolvableEnvTokenError` (a loud exit 2
+ * naming the exact fix) into a dead path the skill discovers at runtime and is
+ * then blamed for.
+ */
+function resolveExecutorEnvForEval(item, ctx, workspaceDir) {
+    if (ctx.perEvalEnv === undefined)
+        return ctx.executorEnv;
+    const declaresFiles = item.entry.files !== undefined && item.entry.files.length > 0;
+    const { input, logged } = ctx.perEvalEnv;
+    const env = resolveDeclaredChildEnv({
+        ...input,
+        ...(declaresFiles ? { workspaceDir } : {}),
+        quiet: logged.done,
+    }).env;
+    logged.done = true;
+    return env;
+}
+/**
+ * Scrub the control arm's env and SAY what it lost.
+ *
+ * The arms are meant to differ in the skill and nothing else, so any var withheld
+ * from one side is a confound. A control arm silently spawned without `PATH` or
+ * without its declared fixtures still runs — it just does worse, which reports as
+ * skill lift. That failure is invisible unless the withholding is announced, and
+ * `dropped` was previously computed and thrown away.
+ */
+function scrubAndReportControlArmEnv(env, harnessRoot, evalId) {
+    // `[]` because vat does not declare model env vars today — `resolveAuth` is
+    // called with no `modelVars`, so none is ever forwarded and the exemption set
+    // has nothing extra to hold. The parameter is REQUIRED rather than defaulted so
+    // that stops being a silent assumption: the day vat forwards a model var, this
+    // line fails to compile and someone decides, instead of the scrub quietly
+    // dropping it and running the control arm on a different model.
+    const { env: scrubbed, droppedForbiddenKey, droppedNamingRoot, retainedLeaks } = scrubControlArmEnv(env, harnessRoot, []);
+    // Two rules, two reasons, two lines. One merged line said "naming the harness
+    // root" about both, which is false for every rule-1 drop — those go by NAME and
+    // their value may name nothing in the harness at all.
+    const withheld = [
+        [droppedForbiddenKey, 'by name, whatever their value'],
+        [droppedNamingRoot, 'because their value names the harness root'],
+    ];
+    for (const [names, why] of withheld) {
+        if (names.length === 0)
+            continue;
+        process.stderr.write(`control arm (${evalId}): withheld ${names.length} env var(s) ${why}: ${names.join(', ')}\n`);
+    }
+    if (retainedLeaks.length > 0) {
+        process.stderr.write(`⚠️  control arm (${evalId}): ${retainedLeaks.join(', ')} name(s) the harness root but are required to run, so they were NOT withheld. ` +
+            `The control arm can read the harness through them — move --out/--workdir off this path before trusting the delta.\n`);
+    }
+    return scrubbed;
+}
+async function gradeOneArm(item, ctx) {
     const evalId = String(item.entry.id);
-    const workspaceDir = resolvePerEvalWorkspaceDir(item.entry, ctx.workspacesRoot);
+    const workspaceDir = resolvePerEvalWorkspaceDir(item.entry, ctx.workspacesRoot, item.arm, ctx.armDirs);
     const onProgress = (chunk) => { process.stderr.write(chunk); };
-    // Resolve the declared `env` against THIS eval's workspace. An eval that
-    // declares no input files has no workspace, so a value using ${fixturesDir}
-    // throws UnresolvableEnvTokenError (exit 2) instead of injecting a dead path.
-    let executorEnv = ctx.executorEnv;
-    if (ctx.perEvalEnv !== undefined) {
-        const { input, logged } = ctx.perEvalEnv;
-        executorEnv = resolveDeclaredChildEnv({
-            ...input,
-            ...(workspaceDir === undefined ? {} : { workspaceDir }),
-            quiet: logged.done,
-        }).env;
-        logged.done = true;
-    }
+    // Resolve the declared `env` against THIS eval's workspace.
+    //
+    // `workspaceDir` is threaded to the token resolver ONLY when the eval actually
+    // declares input `files`, even though every eval now HAS a workspace. The two
+    // are different questions: the workspace is where the executor runs, while
+    // `${fixturesDir}` promises a staged `fixtures/` directory that exists — and
+    // `stageEvalWorkspaces` creates that only for evals declaring files. Passing it
+    // unconditionally silently downgraded `UnresolvableEnvTokenError` (a loud exit 2
+    // naming the exact fix) into a dead path the skill discovers at runtime and gets
+    // blamed for.
+    let executorEnv = resolveExecutorEnvForEval(item, ctx, workspaceDir);
+    // The skill-absent arm is not told where the subject is staged. Handing it that
+    // path would defeat the control outright: the staged dir holds the SKILL.md AND
+    // any executable the skill ships, so a control arm given the path can recover
+    // the entire treatment with one `cat`. Withholding the plugin dir while naming
+    // the staged dir in the prompt was the defect this arm's whole value rests on.
+    const isWithArm = item.arm === 'with';
+    // ...and the same reasoning applied to the ENVIRONMENT, which is the channel the
+    // first round of this fix missed entirely: the run assembles ONE env and both
+    // arms were handed it verbatim, `CLAUDE_PLUGIN_ROOT` included. See
+    // scrubControlArmEnv — prompt, argv, cwd and env are four channels, not three.
+    //
+    // Report what the control arm lost. The arms are supposed to differ in the skill
+    // and nothing else, so any var withheld from one side is a confound the operator
+    // has to be able to see — silently spawning a degraded control is how a harness
+    // manufactures the lift it is meant to measure.
+    if (!isWithArm)
+        executorEnv = scrubAndReportControlArmEnv(executorEnv, ctx.harnessRoot, evalId);
     const outcome = await runExecutorForEval({
         evalId,
         task: item.entry.prompt,
-        subjectStagedDir: ctx.subjectStagedDir,
-        ...(workspaceDir === undefined ? {} : { workspaceDir }),
+        ...(isWithArm ? { subjectStagedDir: ctx.subjectStagedDir } : {}),
+        workspaceDir,
         pluginDirs: item.arm === 'without' ? [] : ctx.pluginDirs,
         env: executorEnv,
         ...(ctx.model === undefined ? {} : { model: ctx.model }),
@@ -777,7 +1114,6 @@ async function runEvalWorker(item, ctx) {
     recordSessionCost(ctx.costAccumulator, outcome.parsed.result?.totalCostUsd);
     // Tool expectations are about the SKILL's tools, so they ride the WITH arm ONLY —
     // the WITHOUT (skill-absent) arm has no skill, hence nothing to judge tools against.
-    const isWithArm = item.arm === 'with';
     const fragment = await runGraderForEval({
         evalId,
         transcript: outcome.transcript,
@@ -802,11 +1138,437 @@ async function runEvalWorker(item, ctx) {
         onProgress,
         costSink: (usd) => recordSessionCost(ctx.costAccumulator, usd),
     });
-    return { ...fragment, arm: item.arm };
+    const graded = withoutGraderContamination(fragment);
+    if (isWithArm)
+        return { ...graded, arm: item.arm };
+    return { ...graded, arm: item.arm, ...baselineContaminationFor(outcome.transcript, ctx, evalId) };
+}
+/**
+ * Run one work item and decide what its failure costs the RUN.
+ *
+ * TREATMENT arm: rethrow. `grading.json` is the treatment arm, the verdict is
+ * computed from it, and a run that cannot produce one has produced nothing — there
+ * is no partial result worth writing, so the throw propagates through `runPipeline`
+ * to `skillTestFailureReason` exactly as before (`ERROR`, reason `internal`).
+ *
+ * CONTROL arm: record and continue. Every failure mode of this arm — executor spawn
+ * error, stall, wall-timeout, grader non-zero exit, grader wrote no fragment,
+ * fragment unparseable twice, nonce mismatch, and a grader that returned a
+ * different number of expectations than the eval declares — used to propagate out of
+ * `runPipeline` and annihilate the treatment work already completed and billed. A
+ * two-eval probe was measured doing precisely that: both treatment executors ran,
+ * both treatment graders ran, the control arm's executor timed out, the run threw,
+ * and `results/` was left holding `provenance.json` and nothing else. The operator
+ * paid for the whole run and got no grading.json, no baseline.json, and an error
+ * that did not even say which arm had died.
+ *
+ * Recording it instead costs the DELTA, which is the only thing the control arm
+ * feeds, and the delta already has a truthful way to be absent (`null`). The
+ * treatment artifacts are written, the failure is named on stderr as it happens and
+ * again in `baselineIntegrity.controlArmFailures`, and `armExpectationSkew` sees an
+ * eval graded on one arm only and withholds the number.
+ *
+ * A {@link RateLimitSignal} is rethrown on BOTH arms — it is not a failure, it is the
+ * pipeline's retry protocol, and swallowing it on the control arm would convert every
+ * transient rate limit into a permanently dead control arm. The worker CANNOT tell a
+ * first 429 from the last one; only the pipeline knows the budget is spent, which is
+ * what {@link rateLimitExhaustedOutcome} is wired to handle.
+ */
+async function runEvalWorker(item, ctx) {
+    try {
+        return { fragment: await gradeOneArm(item, ctx) };
+    }
+    catch (err) {
+        if (err instanceof RateLimitSignal)
+            throw nameTheArm(err, item.arm);
+        if (item.arm === 'with') {
+            if (err instanceof GraderFragmentUnparseableError)
+                return { fragment: unparseableGradingFragment(item, ctx, err) };
+            throw nameTheArm(err, item.arm);
+        }
+        return recordControlFailure(item, err);
+    }
+}
+/**
+ * The TREATMENT arm's outcome for a grader that could not write valid JSON twice
+ * ({@link GraderFragmentUnparseableError}): this eval FAILS, in the same fragment
+ * shape a grader's own failing verdict takes, so it merges, counts and gates
+ * exactly like one — `FAIL n/m`, `FINDINGS` unless `--allow-eval-failure`, the other
+ * evals reported. Every declared expectation is marked failed with the reason as
+ * its `evidence`; a `tool` verdict is NOT invented, because none was produced —
+ * the eval already fails on the output channel, and `tool-eval.json` omitting it
+ * is the truth. Not `InternalHarnessError`: the harness ran the grader twice and
+ * it was the model that could not produce strict JSON.
+ *
+ * Treatment only. On the control arm the same error is recorded as a control
+ * failure (delta withheld) — a synthesized control `fail` would read as lift.
+ */
+function unparseableGradingFragment(item, ctx, err) {
+    process.stderr.write(`\n⚠️  [${ARM_LABEL[item.arm]}] ${err.message}\n` +
+        `   Eval "${err.evalId}" is reported as FAILED (${GRADER_FRAGMENT_UNPARSEABLE}); the run continues.\n`);
+    return {
+        runNonce: ctx.runNonce,
+        evalId: err.evalId,
+        arm: item.arm,
+        expectations: item.entry.expectations.map((text) => ({
+            // Suite text enters an artifact here without a grader's echo in between, so
+            // it takes the same sanitizer a grader's echo would have.
+            text: sanitizeGraderText(text),
+            passed: false,
+            evidence: `${GRADER_FRAGMENT_UNPARSEABLE}: ${err.detail}`,
+        })),
+    };
+}
+/**
+ * The control arm ran out of rate-limit retries — record it and let the run live.
+ *
+ * The retryable case is handled correctly by {@link runEvalWorker} and must stay that
+ * way; this is the EXHAUSTED case, which used to be indistinguishable from it. A
+ * control executor rate-limited on every attempt threw after five retries, taking
+ * down a run in which both treatment executors and both treatment graders had already
+ * completed and been billed, leaving `results/` holding `provenance.json` and nothing
+ * else — the exact failure {@link runEvalWorker} exists to prevent, reached by the one
+ * error class it had to let through.
+ *
+ * The treatment arm still hard-fails: there is no treatment result to salvage, which
+ * is the same asymmetry {@link EvalWorkOutcome} documents.
+ *
+ * Wired as the pipeline's `onRetriesExhausted`, and `--baseline` — which doubles the
+ * spawn count — is the run most likely to need it.
+ */
+function rateLimitExhaustedOutcome(item, error) {
+    if (item.arm === 'with')
+        throw nameTheArm(error, item.arm);
+    return recordControlFailure(item, error);
+}
+/**
+ * Turn one control-arm error into a recorded failure the run carries past. Shared by
+ * the worker's catch and the exhausted-retries handler so both report identically —
+ * an operator must not be able to tell which layer noticed.
+ */
+function recordControlFailure(item, err) {
+    const named = nameTheArm(err, item.arm);
+    // `detail` is schema-required non-empty: an error that stringifies to nothing
+    // would otherwise fail the artifact write and take down the run this whole
+    // branch exists to keep alive.
+    const detail = (named instanceof Error ? named.message : String(named)) ||
+        `[${ARM_LABEL[item.arm]}] failed for eval "${String(item.entry.id)}" with a non-descriptive error`;
+    process.stderr.write(`\n⚠️  ${detail}\n` +
+        '   The control arm for this eval is unusable, so the --baseline delta will be WITHHELD. ' +
+        'The treatment arm is unaffected and the run continues — its grading is what you paid for.\n');
+    return { controlFailure: { evalId: String(item.entry.id), detail } };
+}
+/**
+ * Fragment fields VAT derives itself and therefore never accepts FROM a grader.
+ *
+ * Both are computed from the executor transcript's structure, which the grader is
+ * shown as opaque text and never asked about — so a value in either from a grader
+ * is not merely redundant, it is a claim the grader is in no position to make.
+ */
+const VAT_ATTACHED_FRAGMENT_KEYS = ['contamination', 'degraded'];
+/**
+ * Strip the fields VAT attaches itself, before VAT attaches them.
+ *
+ * The grader's only input is the executor transcript, which untrusted skill code
+ * controls — so a prompt injection there could otherwise fabricate a
+ * `BASELINE CONTAMINATED` verdict carrying attacker-chosen strings into an
+ * operator-facing artifact, or (via `degraded`) claim a scan degraded that did not.
+ * Spreading VAT's value last is NOT sufficient: it is absent on a clean run, so it
+ * overwrites nothing — which is exactly why the omission direction matters more
+ * than the invention one for `degraded`. A grader that omits it cannot re-hide a
+ * degradation VAT detected, because VAT's value is spread after this strip; a
+ * grader that INVENTS it would otherwise stamp a blind-scan warning on a run that
+ * scanned properly, and an operator who learns to ignore that warning stops reading
+ * the real ones.
+ */
+function withoutGraderContamination(fragment) {
+    if (VAT_ATTACHED_FRAGMENT_KEYS.every((key) => fragment[key] === undefined))
+        return fragment;
+    const copy = { ...fragment };
+    for (const key of VAT_ATTACHED_FRAGMENT_KEYS)
+        delete copy[key];
+    return copy;
+}
+/**
+ * How each arm is named in every message that can only be about one of them.
+ *
+ * Spelled out rather than `'with'`/`'without'`: the internal names are a preposition
+ * apart and the operator-facing question is "which half of my experiment died", so
+ * `Executor timed out for eval "e1"` — the actual text a destroyed run used to print
+ * — named neither the arm nor the fact that the OTHER arm had already succeeded and
+ * been thrown away.
+ */
+const ARM_LABEL = {
+    with: 'treatment arm (skill available)',
+    without: 'control arm (skill withheld)',
+};
+/**
+ * Prefix a worker error's message with the arm it came from, IN PLACE.
+ *
+ * Mutating rather than re-wrapping is deliberate. `skillTestFailureReason` reads the
+ * error's own declared reason, and the run's regression test for a forged grader
+ * nonce (`rejects.toThrow(GradingNonceError)`); wrapping every failure in a fresh
+ * `InternalHarnessError` to get a better message would flatten
+ * `EvalFragmentError`/`GradingNonceError`/`RateLimitSignal` into one class and lose
+ * both. The stack is preserved for the same reason.
+ *
+ * Idempotent — a worker error passes through here exactly once, but a retried item
+ * must not accumulate prefixes.
+ */
+function nameTheArm(err, arm) {
+    prefixMessageOnce(err, `[${ARM_LABEL[arm]}] `);
+    return err;
+}
+/**
+ * Lift the run's content needles off the STAGED SKILL.md.
+ *
+ * Staged, not source: the staged copy is byte-identical to what the treatment arm
+ * is given, so a needle taken from it is exactly the text an ambient copy of the
+ * same build would carry.
+ *
+ * The exclusion set is EVERY channel through which vat itself hands the arm text:
+ * the suite's prompts, its `expected_output`, its expectations, and the CONTENTS
+ * of the input `files` fixtures vat stages into the arm's workspace and instructs
+ * it to operate on. An adopter who quotes a sentence of their SKILL.md through any
+ * of the four would otherwise stamp every run contaminated, including the arm that
+ * merely read the input vat gave it — the skill's words arriving THROUGH vat are
+ * not the arm reaching the skill. The fixture channel is the one that was missed,
+ * and it is the worst of the four to get wrong: the summary attached to that
+ * verdict tells the operator to go uninstall an ambient copy of the plugin that
+ * does not exist.
+ *
+ * Fixtures are read from the STAGED workspace rather than the suite directory, for
+ * the same reason the SKILL.md is: that is byte-for-byte what the arm was handed.
+ *
+ * Returns `[]` when there is no staged SKILL.md. A missing SKILL.md is not this
+ * function's error to raise (staging has already validated the subject), and an
+ * unarmed signal is reported as unarmed rather than claimed as clean. A SKILL.md
+ * that is THERE but refused is a different thing — `[]` would silently disarm the
+ * contamination signal for the whole run — so a refusal propagates. The same split
+ * applies to fixtures: see {@link collectFixtureText}.
+ */
+function resolveSkillContentNeedles(subjectStagedDir, evals, 
+/**
+ * Where the staged fixtures live, as the two inputs an arm root is DERIVED from
+ * rather than as the pre-joined path itself — see {@link armWorkspaceRoot}. Taking
+ * a bare `string` here let the call site pass `workspacesRoot` (one level too high,
+ * so every `<id>/` fixture lookup missed) with the whole unit suite still green.
+ *
+ * The `with` arm, because the fixture BYTES are what matter and both arms are
+ * staged from the same source — and `with` is the only arm every run mints.
+ */
+workspaces) {
+    const stagedWorkspaceRoot = armWorkspaceRoot(workspaces.workspacesRoot, workspaces.armDirs, 'with');
+    const skillMdPath = safePath.join(subjectStagedDir, 'SKILL.md');
+    let markdown;
+    try {
+        markdown = readFileSync(skillMdPath, 'utf8');
+    }
+    catch (err) {
+        if (isPathAbsentError(err))
+            return [];
+        throw err;
+    }
+    const excluded = evals
+        .flatMap((entry) => [
+        entry.prompt,
+        entry.expected_output ?? '',
+        ...entry.expectations,
+        ...readStagedFixtures(stagedWorkspaceRoot, entry),
+    ])
+        .join('\n');
+    return skillContentNeedles(markdown, excluded);
+}
+/**
+ * The text of one eval's staged input files. A declared fixture that was never
+ * staged is skipped; one that is there but cannot be read propagates.
+ *
+ * A declared `files[]` entry is legitimately a DIRECTORY — staging does `existsSync`
+ * then a recursive `cpSync` — so this walks one. It used to call `readFileSync` on
+ * the declared path, get `EISDIR`, and skip it: the directory's whole contents were
+ * in the arm's cwd and absent from the exclusion set, and two runs differing ONLY in
+ * whether the same bytes were declared as a file or as their parent directory
+ * reported `contaminated: false` and `contaminated: true` (with a `skill-content`
+ * hit, and a summary telling the operator to uninstall an ambient plugin copy that
+ * does not exist).
+ *
+ * The `catch`'s old claim that skipping "errs the safe way: noisier, never blinder"
+ * is BACKWARDS for this signal. Over-reporting is the harm here, because a false
+ * `contaminated: true` is actioned by discarding a run that was fine. So only an
+ * ABSENT fixture is skipped (staging does `existsSync` then `cpSync`, so a declared
+ * file missing at the source is simply not in the workspace); a fixture that is
+ * there and refused throws, because an exclusion set vat could not compute is a
+ * signal vat cannot stand behind.
+ */
+function readStagedFixtures(stagedWorkspaceRoot, entry) {
+    const texts = [];
+    for (const rel of entry.files ?? []) {
+        const staged = safePath.joinUnderRoot(safePath.joinUnderRoot(stagedWorkspaceRoot, String(entry.id)), rel);
+        collectFixtureText(staged, texts);
+    }
+    return texts;
+}
+/** Depth cap on the fixture walk — a symlink loop must not hang the run's setup. */
+const MAX_FIXTURE_WALK_DEPTH = 32;
+/**
+ * Append the declared fixture's text — the file's, or every file's under it when
+ * it is a directory. `lstat` first, so the ONE absence case (`ENOENT`, or `ENOTDIR`
+ * for a path routed through a file) is the only thing skipped; a symlink at the
+ * declared path is skipped too, for the same reason the walk below skips one.
+ */
+function collectFixtureText(path, into) {
+    let kind;
+    try {
+        kind = lstatSync(path);
+    }
+    catch (err) {
+        if (isPathAbsentError(err))
+            return;
+        throw err;
+    }
+    if (kind.isFile())
+        into.push(readFileSync(path, 'utf8'));
+    else if (kind.isDirectory())
+        walkFixtureDir(path, into, 0);
+}
+/** Append every file's text under `dir`, recursing. Symlinked entries are skipped. */
+function walkFixtureDir(dir, into, depth) {
+    if (depth > MAX_FIXTURE_WALK_DEPTH)
+        return;
+    // `withFileTypes` on the DIRECTORY is one syscall for the whole level, and
+    // `lstat`-shaped: a symlinked entry reports as a link rather than as whatever it
+    // points at, so the walk stays inside the bytes vat actually staged.
+    for (const child of readdirSync(dir, { withFileTypes: true })) {
+        const childPath = safePath.joinUnderRoot(dir, child.name);
+        const kind = direntKind(child);
+        // A link is skipped by design (see above): the walk reads only staged bytes.
+        if (kind === 'symlink')
+            continue;
+        if (kind === 'file')
+            into.push(readFileSync(childPath, 'utf8'));
+        else if (kind === 'directory')
+            walkFixtureDir(childPath, into, depth + 1);
+    }
+}
+/**
+ * Build the detector input for the skill-absent arm. ONE builder, so the hits
+ * written per eval and the `signals` list stamped on the run-level integrity
+ * block are derived from the same values — a second construction site is how the
+ * block would come to claim a detector the scan never actually ran.
+ *
+ * `transcript` and `evalId` are supplied per call; everything else is per-run.
+ *
+ * Takes a NARROWED context rather than the whole {@link EvalRunContext} so the
+ * builder is directly unit-testable — the detector honours `armCwd` (dropping it
+ * there kills three tests) while the wiring that SUPPLIES it was droppable green,
+ * and a dropped `armCwd` anchors every relative climb one directory too high, where
+ * no needle can see it.
+ *
+ * `evalId` is what turns the arm's workspace ROOT into the arm's actual starting
+ * CWD (`<armWorkspaceDir>/<evalId>` — see {@link resolvePerEvalWorkspaceDir}).
+ * The detector needs both and they are one directory apart: containment is judged
+ * against the root (the arm's whole tree is its own), while the cwd walk must
+ * start exactly where the executor did, or every relative climb in the transcript
+ * resolves one level too high and lands where no needle can see it.
+ *
+ * `evalId` is REQUIRED here, and the signals-only caller gets its own entry point
+ * ({@link buildContaminationSignalsInput}) rather than sharing an optional third
+ * argument. An optional parameter is a pin the compiler cannot hold: dropping the
+ * argument at the SCAN call site left all 2270 unit tests green and was caught only
+ * by the integration suite — which this repo's fast implementer loop explicitly says
+ * not to run. That is the same shape as the dropped `armCwd` that anchored the whole
+ * walk one directory too high, also green.
+ */
+function buildContaminationInput(transcript, ctx, evalId) {
+    return contaminationInput(transcript, ctx, evalId);
+}
+/**
+ * The same input built for the "which detectors are ARMED?" question, which is asked
+ * once per run with no transcript and no eval — so it can honestly supply neither,
+ * and `armCwd` is correctly absent rather than accidentally dropped.
+ */
+function buildContaminationSignalsInput(ctx) {
+    return contaminationInput('', ctx, undefined);
+}
+function contaminationInput(transcript, ctx, evalId) {
+    const armRoot = ctx.armDirs.without === undefined || ctx.armDirs.without === ''
+        ? undefined
+        : armWorkspaceRoot(ctx.workspacesRoot, ctx.armDirs, 'without');
+    return {
+        transcript,
+        harnessRoot: ctx.harnessRoot,
+        // Only ever called for the skill-absent arm, so the sibling is `with` — the
+        // treatment arm's live working directory, one `ls ..` away and containing no
+        // harness path for the needles above to find.
+        siblingArmDir: armWorkspaceRoot(ctx.workspacesRoot, ctx.armDirs, 'with'),
+        // The control arm's OWN root, the exact counterpart of the line above. Its
+        // absolute path is stated in the arm's own prompt, so the arm reuses it
+        // constantly; without this the executable-name signal reads every mention of
+        // the arm's own scratch script as a reach into the skill.
+        //
+        // Conditional because this builder also serves `buildContaminationSignalsInput`,
+        // which runs on EVERY run — and a non-baseline run mints no `without` arm, so
+        // `armDirSegment` would (correctly) throw rather than invent a segment.
+        ...(armRoot === undefined ? {} : { armWorkspaceDir: armRoot }),
+        // ...and the eval's own subdirectory under it, which is where the executor was
+        // actually spawned. Absent without an evalId, in which case the walk anchors
+        // at the root — one level high, but the best anchor available.
+        ...(armRoot === undefined || evalId === undefined
+            ? {}
+            : { armCwd: safePath.joinUnderRoot(armRoot, evalId) }),
+        // VAT's private tmp dirs: the held eval suite (the `expected_output` answer
+        // key) and the grader dir (the run's integrity nonce). Both are SIBLINGS of
+        // the arm's cwd under the OS temp dir — two hops or one `$TMPDIR` expansion
+        // away — and contain no harness path, so neither the harness needles nor the
+        // sibling-arm needles can see them. The workspaces root is deliberately NOT
+        // in this list: it is the arm's own legitimate cwd.
+        vatPrivateDirs: [ctx.evalSuiteHoldDir, ctx.graderOutDir],
+        skillContentNeedles: ctx.skillContentNeedles,
+        // The skill-relative PATH (`scripts/csvsum.py`), not the extension-stripped
+        // `name`. An ambient copy of the skill is a COPY, so it reproduces the skill's
+        // internal layout under whatever root it sits at and still matches; the control
+        // arm's own `/tmp/summary.txt` no longer does. `howInvoked` is a command string,
+        // not a path, so it is deliberately NOT used here.
+        ...(ctx.declaredExecutables === undefined
+            ? {}
+            : { executablePaths: ctx.declaredExecutables.map((e) => e.path) }),
+    };
+}
+function baselineContaminationFor(transcript, ctx, evalId) {
+    const { hits, degraded } = detectBaselineContamination(buildContaminationInput(transcript, ctx, evalId));
+    // A degraded scan writes `contaminated: false` with exactly the same bytes as one
+    // that walked the transcript properly, so silence here would be the "quietly wrong
+    // measurement" this whole module exists to prevent. It goes to BOTH surfaces, and
+    // neither substitutes for the other:
+    // stderr reaches the operator watching the run, and the fragment carries it to
+    // `baselineIntegrity.degraded` in baseline.json, which is what a reader has
+    // months later when the stderr scrollback is gone.
+    //
+    // The `evalId` is stamped HERE rather than by the detector: the detector is pure
+    // and scans one transcript with no idea which eval it belongs to, and an
+    // unattributed degradation in a run-level list is unactionable.
+    if (degraded !== undefined) {
+        process.stderr.write(`⚠️  control arm (${evalId}): contamination scan DEGRADED (${degraded.reason}) — ${degraded.detail}. ` +
+            'A clean verdict from this eval is weaker evidence than a clean structured scan: the walk ' +
+            'either stopped resolving paths from this point on, or (for an unparseable transcript) fell ' +
+            'back to flat text matching, which both over- and under-reports.\n');
+    }
+    return {
+        ...(hits.length === 0 ? {} : { contamination: hits }),
+        ...(degraded === undefined ? {} : { degraded: { ...degraded, evalId } }),
+    };
+}
+/**
+ * Which detectors this run armed, for the `signals` field on the integrity block.
+ * Derived from {@link buildContaminationInput} with an empty transcript — the
+ * armed set is a property of the run's paths, never of any one transcript.
+ */
+function contaminationSignalsFor(ctx) {
+    return activeContaminationSignals(buildContaminationSignalsInput(ctx));
 }
 /**
  * Run evals TIER by TIER (ascending), bounded-parallel WITHIN each tier, with a
- * GATE between tiers (issue #145 Phase G). After a tier completes, apply the
+ * GATE between tiers. After a tier completes, apply the
  * default gate policy ({@link shouldGateAfterTier}) over that tier's WITH-arm
  * fragments: if any eval in the tier did not fully pass, do NOT launch the higher
  * (more expensive) tiers — their evals are recorded as SKIPPED (a distinct state,
@@ -819,9 +1581,16 @@ async function runEvalWorker(item, ctx) {
 async function runEvalsTiered(input) {
     const groups = groupEvalsByTier(input.evals);
     const fragments = [];
+    const controlFailures = [];
     for (const [index, group] of groups.entries()) {
-        const tierFragments = await input.runTier(buildEvalWorkItems(group.evals, input.baseline));
+        const outcomes = await input.runTier(buildEvalWorkItems(group.evals, input.baseline));
+        const tierFragments = outcomes.flatMap((o) => (o.fragment === undefined ? [] : [o.fragment]));
+        controlFailures.push(...outcomes.flatMap((o) => (o.controlFailure === undefined ? [] : [o.controlFailure])));
         fragments.push(...tierFragments);
+        // A control-arm failure deliberately does NOT gate. The gate asks whether the
+        // SKILL failed a cheap expectation before spending on expensive ones, and the
+        // control arm is not the skill — treating its executor timeout as a gating
+        // failure would stop a run whose treatment tier passed everything.
         const withArm = tierFragments.filter((f) => f.arm !== 'without');
         if (!shouldGateAfterTier(withArm))
             continue;
@@ -832,14 +1601,14 @@ async function runEvalsTiered(input) {
         // Legibility (required): name the skipped tiers on stderr so a fail-fast run is
         // never mistaken for a smaller passing suite. stdout stays machine-readable.
         process.stderr.write(formatSkippedTiersSummary(skipped) + '\n');
-        return { fragments, skipped };
+        return { fragments, controlFailures, skipped };
     }
-    return { fragments };
+    return { fragments, controlFailures };
 }
 /** Resolve the run's grading/friction/baseline/tool-eval artifact paths under
  *  `resultsDir`. Single source of truth for the filenames (used by the
  *  pre-pipeline stale wipe AND the post-merge writer). */
-export function resolveArtifactPaths(resultsDir) {
+function resolveArtifactPaths(resultsDir) {
     return {
         gradingOut: safePath.join(resultsDir, 'grading.json'),
         frictionOut: safePath.join(resultsDir, 'friction.json'),
@@ -856,11 +1625,52 @@ export function resolveArtifactPaths(resultsDir) {
  * run's. Wiping all three up front closes that cross-run leak. Best-effort
  * (`force: true`) — a missing file is fine.
  */
-export function wipeStaleArtifacts(paths) {
-    rmSync(paths.gradingOut, { force: true });
-    rmSync(paths.frictionOut, { force: true });
-    rmSync(paths.baselineOut, { force: true });
-    rmSync(paths.toolEvalOut, { force: true });
+function wipeStaleArtifacts(paths) {
+    for (const path of [paths.gradingOut, paths.frictionOut, paths.baselineOut, paths.toolEvalOut]) {
+        rmSync(path, { force: true });
+        // ...and the quarantined copy a PRIOR run's failed gate set aside. It is just as
+        // stale, and left alone `results/` accumulates one `.rejected` per broken run
+        // with nothing to say which is the newest.
+        rmSync(rejectedArtifactPath(path), { force: true });
+    }
+}
+/**
+ * Create `results/` and write this run's provenance — unless this is a DRY RUN, in
+ * which case it does nothing at all.
+ *
+ * NOTHING under `results/` may be created, written, or removed by a dry run. The
+ * harness root is a deterministic function of the subject, so `results/` here holds
+ * the PREVIOUS run's output: creating the dir, rewriting `provenance.json`, and (at
+ * the call site) wiping the stale artifacts meant "what would this cost next time?"
+ * destroyed the run the operator was about to read. Measured twice: a real
+ * `--baseline` run left `[baseline, friction, grading, provenance, tool-eval].json`
+ * and a subsequent `--dry-run` against the same subject left `[provenance.json]`.
+ *
+ * A function rather than an `if` at the call site because the orchestrator is at its
+ * cognitive-complexity ceiling; the RULE is what matters and it is written here.
+ */
+function openResultsDir(input) {
+    if (input.dryRun)
+        return;
+    // 0700, matching the harness root. `grading.json` quotes the executor transcript
+    // verbatim in each expectation's evidence, so whatever the skill read out of its
+    // input files ends up here as text — and unlike the workspaces, results/ SURVIVES
+    // `--keep` by design.
+    mkdirSyncReal(input.resultsDir, { recursive: true, mode: 0o700 });
+    writeFileSync(input.provenancePath, JSON.stringify(input.provenance, null, 2) + '\n', 'utf-8');
+}
+/**
+ * Where a results artifact goes when it fails the fail-closed gate.
+ *
+ * The gate throws and `results/` is deliberately RETAINED, so the file that failed
+ * used to sit there under its authoritative name — the one artifact whose schema
+ * violation is BY DEFINITION a harness bug, indistinguishable in a CI archive from a
+ * good one until the next run's wipe. Renamed rather than deleted: the bytes are what
+ * a maintainer needs in order to see which refine tripped, and a name ending in
+ * `.rejected` cannot be mistaken for the artifact a reader came for.
+ */
+function rejectedArtifactPath(path) {
+    return `${path}.rejected`;
 }
 /**
  * Merge the run's grader fragments and WRITE the run artifacts (vat is the SOLE
@@ -874,49 +1684,253 @@ export function wipeStaleArtifacts(paths) {
  * can compute the COMPOSITE verdict. Every fragment's per-run nonce is re-verified
  * inside {@link mergeFragmentsToGrading}.
  */
-function writeRunArtifactsAndReconcile(fragments, runNonce, paths) {
+/**
+ * Collect per-eval baseline-integrity findings from the WITHOUT-arm fragments.
+ * Fragments with no `contamination` are clean and contribute nothing. Pure.
+ */
+/**
+ * What one arm graded, per eval — the SINGLE derivation behind both the parity check
+ * ({@link armExpectationSkew}, which reads `total`) and the A/B delta (which reads
+ * `passed`). Deriving those separately would let the two blocks in `baseline.json`
+ * disagree about the same arm. Pure.
+ *
+ * Exported for its own unit test, not for reuse. Setting `passed` to `total` here
+ * failed zero unit tests and three integration ones: every unit fixture graded
+ * everything it declared, so the two fields were numerically identical in all of
+ * them and the ONE thing this function exists to keep distinct was unpinned.
+ */
+function gradedCounts(fragments) {
+    return fragments.map((f) => ({
+        evalId: f.evalId,
+        passed: f.expectations.filter((e) => e.passed).length,
+        total: f.expectations.length,
+    }));
+}
+/**
+ * Both per-eval integrity channels the WITHOUT-arm fragments carry, lifted in ONE
+ * pass: what the scan FOUND (`findings`) and how well it was able to look
+ * (`degraded`).
+ *
+ * Together, not separately, because they are read together: `contaminated: false`
+ * means "looked properly and found nothing" only when `degraded` is empty, and a
+ * caller that collects one and forgets the other publishes the first half of that
+ * sentence as if it were the whole of it. That is not hypothetical — the degradation
+ * was detected, schema'd, summarised and rendered, and then never collected, so
+ * every `baseline.json` shipped claiming a clean scan it had not necessarily run.
+ */
+function collectBaselineFindings(withoutArm) {
+    return {
+        findings: withoutArm
+            .filter((f) => f.contamination !== undefined && f.contamination.length > 0)
+            .map((f) => ({ evalId: f.evalId, hits: f.contamination })),
+        degraded: withoutArm.flatMap((f) => (f.degraded === undefined ? [] : [f.degraded])),
+    };
+}
+/**
+ * Restate a {@link SkippedEvalsSummary} as the delta block's truncation record.
+ *
+ * A separate shape rather than embedding the tier summary wholesale: the delta cares
+ * about WHAT IS MISSING FROM ITS TOTALS, which is the flat set of evals that never
+ * ran on either arm, while `SkippedEvalsSummary` is organised for the tier reporter.
+ * Pure.
+ */
+function deltaTruncationFor(skipped) {
+    if (skipped === undefined || skipped.totalSkipped === 0)
+        return null;
+    return {
+        gatedByTier: skipped.gatedByTier,
+        firstSkippedTier: skipped.firstSkippedTier,
+        totalSkipped: skipped.totalSkipped,
+        evalIds: skipped.tiers.flatMap((tier) => tier.evalIds),
+    };
+}
+/**
+ * The stderr report for a `--baseline` run: the delta line, then the ⚠️ banner when
+ * the run is unusable. Pure — the caller decides WHEN it goes out, which is the
+ * whole point (see the call site).
+ *
+ * Composed on EVERY baseline run, including a withheld one: "the delta is
+ * unavailable, and here is why" is the report an operator needs, and a run that
+ * silently prints nothing is indistinguishable from a vat that still does not
+ * compute a delta at all. The delta line comes FIRST so an unusable run reads as the
+ * number followed by the caveat that explains it — the banner is the only line that
+ * says the number means nothing, so it must never be the one scrolled away.
+ *
+ * ⚠️ IT IS NOT LITERALLY LAST ON THE SCREEN, and this docblock used to say it was.
+ * The CLI prints `Harness:`, `Results:`, `Workspaces:` and `Summary:` AFTER the
+ * harness returns, so 2–4 lines always follow the banner. What this ordering buys is
+ * that the banner is not separated from its delta by up to ~1250 rows of friction —
+ * a bounded, short tail after it, not none. Moving those CLI lines above the harness
+ * report is the CLI's call, not this module's.
+ *
+ * THREE conditions have to hold before the caveat is dropped, not two. A DEGRADED
+ * scan is neither contaminated nor incomparable, so a two-clause gate printed
+ * `Baseline delta: +0 (with skill: 2/2, without skill: 2/2).` and exited 0 while
+ * `baseline.json` carried `degraded: [...]` and an `integrity.summary` ending in a
+ * `⚠️ DEGRADED SCAN: …` sentence that was composed, written to the artifact, and
+ * never emitted anywhere. Everything this module says about telling "checked and clean"
+ * apart from "checked with the blunt instrument" was true of the artifact and false
+ * of the terminal — and the terminal is the only surface an operator reads.
+ *
+ * A CONTAMINATED run still reports its number, deliberately. Contamination does not
+ * make the subtraction illegal — both arms were graded to the same depth, so the
+ * arithmetic is sound; what it invalidates is INTERPRETING that number as skill
+ * lift, which is precisely what the ⚠️ line says ("The reported delta is NOT a
+ * measure of skill lift"), and what `vat-skill-testing.md` turns into the
+ * instruction "discard the delta". Withholding it here would (a) leave that sentence
+ * referring to a number that was never reported, and (b) spend `null` — which means
+ * "these arms cannot be subtracted", a different failure with a different remedy —
+ * on a run where they can be. A contaminated delta is also diagnostic in its own
+ * right: it collapses toward zero, which is corroboration, not noise.
+ */
+function formatBaselineReport(delta, integrity) {
+    const line = `\n${formatBaselineDeltaLine(delta)}\n`;
+    // Any of the three makes the delta unusable-as-stated, and all three are only
+    // visible on stderr — nobody opens baseline.json unprompted.
+    if (!integrity.contaminated && integrity.comparable && integrity.degraded.length === 0)
+        return line;
+    return `${line}\n⚠️  ${integrity.summary}\n\n`;
+}
+const IN_PLACE_SUBJECT_FRICTION = {
+    severity: 'low',
+    category: 'path-assumption',
+    message: 'In-place skill (publish: false) — staged from source; links that leave the skill directory are not ' +
+        'present in the harness, so an eval that depends on them tests the skill without them.',
+};
+function writeRunArtifactsAndReconcile(input) {
+    const { fragments, runNonce, paths, signals, controlFailures } = input;
     const { withArm, withoutArm } = partitionFragmentsByArm(fragments);
-    const grading = mergeFragmentsToGrading(withArm, runNonce);
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
+    let baselineReport;
+    const grading = mergeFragmentsToGrading(withArm, runNonce, 'with');
     writeFileSync(paths.gradingOut, JSON.stringify(grading, null, 2) + '\n', 'utf-8');
-    const friction = mergeFragmentsToFriction(fragments);
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
+    const merged = mergeFragmentsToFriction(fragments);
+    const friction = input.subjectInPlace ? { items: [IN_PLACE_SUBJECT_FRICTION, ...merged.items] } : merged;
     writeFileSync(paths.frictionOut, JSON.stringify(friction, null, 2) + '\n', 'utf-8');
     // Tool verdicts come from the WITH arm ONLY — the WITHOUT/skill-absent arm never
     // carries toolExpectations, so its fragments have no `tool` body to merge.
     const toolEval = mergeFragmentsToToolEval(withArm);
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
     writeFileSync(paths.toolEvalOut, JSON.stringify(toolEval, null, 2) + '\n', 'utf-8');
-    if (withoutArm.length > 0) {
-        const baseline = mergeFragmentsToGrading(withoutArm, runNonce);
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
-        writeFileSync(paths.baselineOut, JSON.stringify(baseline, null, 2) + '\n', 'utf-8');
+    // `controlFailures` is the second disjunct, and it is load-bearing: when EVERY
+    // control eval died there are no WITHOUT-arm fragments at all, and gating on the
+    // fragment count alone would write no baseline.json — so the one run that most
+    // needs to say "the control arm did not happen" would say nothing, exactly like a
+    // run where `--baseline` was never passed. `mergeFragmentsToGrading` over an empty
+    // arm yields a zero/zero report, which is the honest shape for it.
+    if (withoutArm.length > 0 || controlFailures.length > 0) {
+        const baseline = mergeFragmentsToGrading(withoutArm, runNonce, 'without');
+        // Stamp the integrity verdict onto baseline.json (GradingReportSchema is
+        // .passthrough(), so the extra field is contract-legal). Written on EVERY
+        // baseline run, clean or not: a reader must be able to tell "checked and
+        // clean" from "produced before this check existed", and only an
+        // unconditional field does that.
+        // Parity BEFORE the verdict prose, because it decides whether the number the
+        // prose is about means anything. Deliberately NOT a throw: the control arm's
+        // grader misbehaving must not discard a perfectly good treatment result — it
+        // must make the DELTA say, in writing, that it cannot be subtracted.
+        // ONE derivation per arm, feeding BOTH the parity check and the delta. The
+        // whole reason `ArmEvalGrade` carries `passed` alongside `total` is that these
+        // two blocks must provably be talking about the same numbers — re-deriving the
+        // counts for the delta would let `baselineIntegrity.skew` and `baselineDelta`
+        // disagree about one run, and a reader holding two contradicting blocks has no
+        // way to tell which one to believe.
+        const withGrades = gradedCounts(withArm);
+        const withoutGrades = gradedCounts(withoutArm);
+        const skew = armExpectationSkew(withGrades, withoutGrades);
+        const { findings, degraded } = collectBaselineFindings(withoutArm);
+        const integrity = summarizeBaselineIntegrity({
+            findings,
+            // WHAT WAS SCANNED, not what was armed. `signals` is derived from the run's
+            // PATHS, so it is fully populated on a run where every control eval died before
+            // producing a transcript — and the shipped documentation teaches `signals` as
+            // exactly the discriminator between a clean verdict and a blind one ("an empty
+            // list means nothing was looking"). A control-timeout run wrote
+            // `"contaminated": false` alongside `checked by: harness-path, sibling-arm,
+            // vat-private-dir, skill-content` having scanned zero transcripts: the list was
+            // full and nothing was looking. With no control fragment there is no
+            // observation, so there is no detector to claim.
+            signals: withoutArm.length === 0 ? [] : signals,
+            // The same fact `signals` is gated on, stated as a number rather than inferred
+            // from an empty list — it gates the VERDICT PROSE, which `signals` cannot reach.
+            // Without it the summary opened "No skill-absent eval was observed reaching the
+            // skill. The A/B delta is interpretable as instruction lift", which is a positive
+            // claim about an observation that never happened, with the correction third.
+            // Required, not defaulted, for the same reason `signals` is: a default lets every
+            // caller overclaim in the "clean" direction, which is where it gets believed.
+            // Control evals whose transcript actually reached the scanner — NOT evals declared.
+            observedEvals: withoutArm.length,
+            skew,
+            degraded,
+            controlArmFailures: controlFailures,
+        });
+        // The SUBTRACTION `--baseline` exists to perform, and until this it was never
+        // performed anywhere: vat wrote two same-shaped artifacts and left the operator
+        // to do the arithmetic by hand while the docs said it "reports the delta".
+        // `skew` is passed rather than re-checked so the withholding decision has
+        // exactly one author.
+        // `truncated` rides along rather than withholding: fail-fast stops BOTH arms at
+        // the same tier boundary, so the arithmetic over the tiers that ran is sound —
+        // what it changes is the SCOPE of the claim. Without it, a tier-0 failure in an
+        // 8-eval suite printed `+1 (with skill: 2/2, without skill: 1/2)`, which is
+        // byte-identical to a complete 2-eval suite. And this is the COMMON case, not a
+        // corner: `--baseline` exists to measure a skill that may not be helping, so a
+        // WITH-arm failure — the thing that fires the gate — is the expected state of an
+        // interesting baseline run.
+        const delta = computeBaselineDelta({
+            withArm: withGrades,
+            withoutArm: withoutGrades,
+            skew,
+            controlArmFailures: controlFailures,
+            truncated: deltaTruncationFor(input.skipped),
+        });
+        // A SIBLING of `baselineIntegrity`, not a field inside it, because they answer
+        // different questions: integrity answers "may I subtract these two arms?", and
+        // the delta IS the subtraction. Burying the number inside the block that
+        // qualifies it would also mean a reader who wants only the lift has to
+        // understand the contamination vocabulary first. (GradingReportSchema is
+        // .passthrough(), so both extra keys are contract-legal — see the comment above.)
+        writeFileSync(paths.baselineOut, JSON.stringify({ ...baseline, baselineIntegrity: integrity, baselineDelta: delta }, null, 2) + '\n', 'utf-8');
+        // RETURNED, not written. The delta line and the ⚠️ banner must not be separated
+        // from each other by the friction report: `emitFrictionReport` runs from the
+        // harness `finally`, long after this point, and its two caps (50 lines × 2000
+        // chars) allow 101,245 characters — about 1250 rows of an 80-column terminal —
+        // between the banner and the prompt. The banner is the only line that says the
+        // number above it is meaningless, so pushing it off-screen is the same failure as
+        // not printing it. Handing the text back lets the caller emit friction FIRST and
+        // this after it; a total-output budget on friction would only shorten the
+        // distance, not end it. (The banner is not the LAST line of the process's output
+        // — the CLI adds `Harness:`/`Results:`/`Workspaces:`/`Summary:` after the harness
+        // returns. That tail is 2–4 lines and bounded, which is the point.)
+        baselineReport = formatBaselineReport(delta, integrity);
     }
-    return { verdict: reconcileGrading(grading), toolEval };
+    return {
+        verdict: reconcileGrading(grading),
+        toolEval,
+        ...(baselineReport === undefined ? {} : { baselineReport }),
+    };
 }
 /**
- * The COMPOSITE run verdict (issue #145 Phase T): the run passes only when BOTH
+ * The COMPOSITE run verdict: the run passes only when BOTH
  * the prose-expectation grading passed AND every tool-expectation verdict passed.
  * Tool verdicts live in tool-eval.json (a SEPARATE channel — C2); this combines
  * the two at the exit-code layer WITHOUT mixing the channels' data. When no eval
  * declared `toolExpectations`, `toolEval.evals` is empty and this equals
  * `outputAllPassed`. Pure + unit-testable.
  */
-export function computeCompositeVerdict(outputAllPassed, toolEval) {
+function computeCompositeVerdict(outputAllPassed, toolEval) {
     return outputAllPassed && toolEval.evals.every((v) => v.passed);
 }
 /**
  * The run's final pass/fail after cost-tiered fail-fast: the {@link
  * computeCompositeVerdict} of the tiers that RAN, AND no tiers were skipped. A
  * fail-fast run that gated higher tiers is NEVER a pass (skipped ≠ passed) — this
- * forces `false` so the exit code is EvalFailure (4), never downgraded to 0 by the
+ * forces `false` so the exit code is FINDINGS, never downgraded to 0 by the
  * composite path alone. Pure + unit-testable.
  */
-export function resolveCompositeAllPassed(outputAllPassed, toolEval, skipped) {
+function resolveCompositeAllPassed(outputAllPassed, toolEval, skipped) {
     return computeCompositeVerdict(outputAllPassed, toolEval) && skipped === undefined;
 }
 /** Fold one session's `total_cost_usd` into the accumulator; a non-number (mock spawn / missing result) is ignored. */
-export function recordSessionCost(acc, totalCostUsd) {
+function recordSessionCost(acc, totalCostUsd) {
     if (typeof totalCostUsd !== 'number' || !Number.isFinite(totalCostUsd))
         return;
     acc.totalUsd += totalCostUsd;
@@ -927,7 +1941,7 @@ export function recordSessionCost(acc, totalCostUsd) {
  * when no session reported a cost (e.g. every spawn was a test mock) so the suffix
  * never adds noise to a run with no cost signal. Pure + unit-testable.
  */
-export function formatRunCostSuffix(cost) {
+function formatRunCostSuffix(cost) {
     if (cost === undefined || cost.sessions === 0)
         return '';
     return ` | ≈$${cost.totalUsd.toFixed(2)} across ${cost.sessions} session${cost.sessions === 1 ? '' : 's'}`;
@@ -939,7 +1953,7 @@ export function formatRunCostSuffix(cost) {
  * verdict failed, a `(N tool)` suffix names how many — so a composite FAIL whose
  * OUTPUT counts look all-green (e.g. `FAIL 3/3 (1 tool)`) is self-explaining. Pure.
  */
-export function buildRunSummary(verdict, toolEval, compositeAllPassed) {
+function buildRunSummary(verdict, toolEval, compositeAllPassed) {
     const base = `${compositeAllPassed ? 'PASS' : 'FAIL'} ${verdict.passed}/${verdict.total}`;
     const toolFailures = toolEval.evals.filter((v) => !v.passed).length;
     return toolFailures > 0 ? `${base} (${toolFailures} tool)` : base;
@@ -951,7 +1965,7 @@ export function buildRunSummary(verdict, toolEval, compositeAllPassed) {
  * tiers ALWAYS reads FAIL (skipped ≠ passed forces `compositeAllPassed` false at
  * the call site), so the base line is already FAIL when the note is present.
  */
-export function buildRunSummaryWithSkips(verdict, toolEval, compositeAllPassed, skipped, cost) {
+function buildRunSummaryWithSkips(verdict, toolEval, compositeAllPassed, skipped, cost) {
     // The spend suffix rides the verdict line (like `(N tool)`); the skipped-tiers
     // note stays on its own following line.
     const base = buildRunSummary(verdict, toolEval, compositeAllPassed) + formatRunCostSuffix(cost);
@@ -965,35 +1979,88 @@ export function buildRunSummaryWithSkips(verdict, toolEval, compositeAllPassed,
  * safe precisely because vat always writes these after the merge succeeds.
  */
 function assertVatWroteArtifact(path, validate, label) {
-    // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
     if (!existsSync(path)) {
         throw new InternalHarnessError(`vat did not write ${label} at ${path} after the merge — harness bug.`);
     }
     let raw;
     try {
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
         raw = JSON.parse(readFileSync(path, 'utf-8'));
     }
     catch (err) {
-        throw new InternalHarnessError(`vat-written ${label} at ${path} is not valid JSON (harness bug): ${err instanceof Error ? err.message : String(err)}`);
+        throw new InternalHarnessError(`vat-written ${label} at ${path} is not valid JSON (harness bug)${quarantineArtifact(path)}: ` +
+            `${err instanceof Error ? err.message : String(err)}`);
     }
     try {
         validate(raw);
     }
     catch (err) {
-        throw new InternalHarnessError(`vat-written ${label} at ${path} failed its schema (harness bug): ${err instanceof Error ? err.message : String(err)}`);
+        throw new InternalHarnessError(`vat-written ${label} at ${path} failed its schema (harness bug)${quarantineArtifact(path)}: ` +
+            `${err instanceof Error ? err.message : String(err)}`);
+    }
+}
+/**
+ * Move a gate-failing artifact aside and say where it went. See
+ * {@link rejectedArtifactPath}.
+ *
+ * Best-effort and never throws: this runs on the way OUT of a run that is already
+ * failing, and a cleanup that replaces the real diagnosis with a rename error would
+ * be strictly worse than a file left in place. If the rename fails the file is
+ * removed instead — an absent artifact reads as "vat wrote nothing", which is at
+ * least not a false claim of authority. If THAT fails too, the file is still there
+ * under its authoritative name, and the returned clause says so with both errors:
+ * the operator is about to read a CI archive, and a bare '' told them nothing.
+ */
+function quarantineArtifact(path) {
+    const rejected = rejectedArtifactPath(path);
+    let renameFailure;
+    try {
+        renameSync(path, rejected);
+        return `; moved aside to ${rejected} so it is not read as authoritative`;
+    }
+    catch (err) {
+        renameFailure = err instanceof Error ? err.message : String(err);
+    }
+    try {
+        rmSync(path, { force: true });
+        return '; removed so it is not read as authoritative';
+    }
+    catch (err) {
+        return (`; it could not be moved aside (${renameFailure}) or removed ` +
+            `(${err instanceof Error ? err.message : String(err)}), so it is STILL THERE under its ` +
+            'authoritative name — do not read it as authoritative');
     }
 }
 /**
  * Post-merge D2 gate over the vat-written results artifacts (vat is SOLE writer):
- * grading.json, friction.json, and tool-eval.json must each exist + parse + validate.
- * A separate, explicit check (NOT folded into any prompt-invariant linter) — fail-closed
- * and safe because vat wrote every one of these itself just above.
+ * grading.json, friction.json, tool-eval.json — and, on a baseline run, baseline.json
+ * — must each exist + parse + validate. A separate, explicit check (NOT folded into
+ * any prompt-invariant linter) — fail-closed and safe because vat wrote every one of
+ * these itself just above.
+ *
+ * `baseline.json` was EXEMPT from this gate for as long as the gate has existed, and
+ * it is the artifact the exemption fits worst: on a `--baseline` run it is the only
+ * durable record of the thing the command sells, and this branch has already been
+ * bitten once by a default run deleting `results/` out from under it. Its validation
+ * is also the strictest of the four — {@link BaselineArtifactSchema} carries the
+ * `passed <= total` and `delta === with − without` refines, so a merge that produced
+ * a self-contradicting number fails the run here instead of being written down.
+ *
+ * `baselineRun` gates it rather than "does the file exist", deliberately: an
+ * existence check would pass on the one failure that matters (vat asked for a
+ * baseline and wrote nothing), which is exactly what a fail-CLOSED gate must not do.
+ * A non-baseline run writes no baseline.json at all and is not asked about one.
  */
-function assertVatWroteArtifacts(paths) {
+function assertVatWroteArtifacts(paths, baselineRun) {
     assertVatWroteArtifact(paths.gradingOut, (raw) => { GradingReportSchema.parse(raw); }, 'grading.json');
     assertVatWroteArtifact(paths.frictionOut, (raw) => { FrictionReportSchema.parse(raw); }, 'friction.json');
     assertVatWroteArtifact(paths.toolEvalOut, (raw) => { ToolEvalReportSchema.parse(raw); }, 'tool-eval.json');
+    if (!baselineRun)
+        return;
+    assertVatWroteArtifact(paths.baselineOut, (raw) => {
+        // Both halves: the grading report it IS, and the two blocks vat stamps onto it.
+        GradingReportSchema.parse(raw);
+        BaselineArtifactSchema.parse(raw);
+    }, 'baseline.json');
 }
 // ---------------------------------------------------------------------------
 // Main orchestrator
@@ -1002,17 +2069,56 @@ function assertVatWroteArtifacts(paths) {
  * Domain orchestrator for `vat skill test run`. Pure of process.exit — all
  * exit-code decisions live in the caller (run.ts).
  */
-export async function runSkillTestHarness(opts) {
+/**
+ * Resolve WHERE this run lives, and whether the harness owns that location.
+ *
+ * `--out` names the harness root exactly; `--workdir` names the BASE the root is
+ * derived under. They are mutually exclusive: passing both used to silently
+ * discard `--workdir` — the run went entirely to `--out` and the `--workdir` path
+ * was never even created — so an operator using the pair to separate the
+ * executor's cwd from the staged trees got no separation and no warning.
+ *
+ * `harnessCreated` is true only for the derived-under-OS-tmp case; an explicit
+ * `--out` or `--workdir` is a location the user owns and is never auto-removed.
+ */
+function resolveHarnessLocation(opts) {
+    if (opts.out !== undefined && opts.workdir !== undefined) {
+        throw new HarnessLocationError('--out and --workdir are mutually exclusive: --out names the harness root exactly, ' +
+            '--workdir names the base it is derived under. Pass one.');
+    }
     // §7 workdir safety: refuse a --workdir whose ancestry contains CLAUDE.md/.claude
     // BEFORE deriving the harness root from it (defense in depth with --setting-sources "").
     if (opts.workdir !== undefined) {
-        assertSafeWorkdir(opts.workdir);
+        // Bounded at the user's home directory, which the walk stops BEFORE inspecting.
+        //
+        // Unbounded, this refuses every `--workdir` on Windows for every Claude Code
+        // user: the OS temp dir there lives INSIDE the profile
+        // (`C:\Users\<name>\AppData\Local\Temp`), so the ancestry walk climbs out of tmp
+        // and finds the ambient `~/.claude` — and the resulting message, "Use an OS-tmp
+        // location", names the thing the user just did. Unsatisfiable advice. It cannot
+        // reproduce on POSIX, where tmp is never under `$HOME`, nor on the GitHub Windows
+        // runner, whose `runneradmin` profile has no `~/.claude`; it needs a real Windows
+        // developer machine, which is where it was found.
+        //
+        // Not a relaxation of the threat model: the gate refuses a workdir inside a
+        // PROJECT, and `$HOME/.claude` is global config that every Claude Code user has on
+        // every platform — it was never the target. A project at `$HOME/proj` is still
+        // caught, because the walk inspects `$HOME/proj` and stops only at `$HOME` itself.
+        assertSafeWorkdir(opts.workdir, homedir());
     }
-    const harnessRoot = opts.out ?? resolveHarnessRoot([opts.subject], opts.workdir);
-    // Only auto-remove the dir the harness itself created under OS tmp. An explicit
-    // --out (exact dir) or --workdir (user-chosen base) is a location the user owns;
-    // treat it like --keep and never delete it.
-    const harnessCreated = opts.out === undefined && opts.workdir === undefined;
+    return {
+        // `--out` is RESOLVED to an absolute path, not taken verbatim. A relative
+        // `--out ./h` used to be stored as `./h` and then used as a contamination
+        // needle — a two-character string that matches almost any transcript, so every
+        // eval reported contaminated and the operator was told to discard a clean run.
+        harnessRoot: opts.out === undefined
+            ? resolveHarnessRoot([opts.subject], opts.workdir)
+            : safePath.resolve(opts.out),
+        harnessCreated: opts.out === undefined && opts.workdir === undefined,
+    };
+}
+export async function runSkillTestHarness(opts) {
+    const { harnessRoot, harnessCreated } = resolveHarnessLocation(opts);
     const repoRoot = opts.repoRoot ?? harnessRoot;
     // Two distinct questions, deliberately not one value. `evalsRef` is what the
     // adopter ASKED FOR — a `test.evals`/`--evals` value, or `undefined` for the
@@ -1055,6 +2161,31 @@ export async function runSkillTestHarness(opts) {
     // when the subject carried no suite, which is the common case) so the staging call
     // stays branch-free; removed by the same cleanup as the grader dir.
     const evalSuiteHoldDir = resolveEvalSuiteHoldDir(randomBytes(16).toString('hex'));
+    // The executor's per-eval working directories. Like the grader and hold dirs
+    // these live OUTSIDE harnessRoot — see resolveWorkspacesRoot for why that
+    // separation is load-bearing for --baseline — so they need their own cleanup.
+    const workspacesRoot = resolveWorkspacesRoot(randomBytes(16).toString('hex'));
+    // Whether those workspaces OUTLIVE the run. `--keep` is the ONLY thing that
+    // retains them.
+    //
+    // Deliberately NOT mirrored on `cleanupHarness`'s rule (keep OR user-owned
+    // location). That rule exists because `--out`/`--workdir` name a directory the
+    // USER owns, so vat has no business deleting it. Since these moved out of the
+    // harness root they live under OS tmp, which the user did not choose and does
+    // not manage — carrying the rule across would orphan a
+    // `vat-skill-test-ws-<token>` dir on every `--out` run, forever. The location
+    // changed, so the retention policy has to change with it.
+    //
+    // Decided ONCE, here, because two consumers ask the question — cleanup (delete
+    // or not) and the returned `workspacesPath` (report or not) — and a second copy
+    // of the predicate would drift, which is exactly how the result came to name a
+    // directory cleanup had already removed.
+    const retainWorkspaces = opts.keep === true;
+    // The opaque per-arm segments beneath it. Minted here, with the roots, so the
+    // same values reach staging, the executor's cwd, and the contamination
+    // detector's sibling-arm needle — one source of truth for a value whose whole
+    // job is to be unguessable and to carry no meaning.
+    const armDirs = mintArmWorkspaceDirs(opts.baseline === true);
     const cleanup = () => {
         // Reap any still-in-flight executor/grader children FIRST. On the concurrent
         // error path `Promise.all` rejects without cancelling its siblings, and a
@@ -1064,7 +2195,13 @@ export async function runSkillTestHarness(opts) {
         // NOTE: the registry is a module-level singleton, so this assumes one run per
         // process (true for the CLI). A library embedding two concurrent runs in one
         // process would need per-run child scoping instead.
-        killAllActiveClaudeChildren();
+        //
+        // `cleanupHarness`, `removeVatOnlyDir` and `lock.release()` each swallow
+        // explicitly at their own definition, for the stated reason that cleanup "must
+        // not mask the run's real outcome". `killAllActiveClaudeChildren` lives in
+        // another package, so its guard is applied here — see {@link swallowCleanupFailure}
+        // for why a call-site guard is the weaker of the two forms.
+        swallowCleanupFailure(killAllActiveClaudeChildren);
         lock.release();
         cleanupHarness(harnessRoot, { keep: opts.keep === true, created: harnessCreated });
         removeVatOnlyDir(graderOutDir);
@@ -1072,6 +2209,10 @@ export async function runSkillTestHarness(opts) {
         // --keep, which retains the harness dir for inspection but has no business
         // retaining the key), and never conditionally.
         removeVatOnlyDir(evalSuiteHoldDir);
+        // Per-eval workspaces are the executor's cwd; `retainWorkspaces` (minted with
+        // the root above) is the single author of whether they survive.
+        if (!retainWorkspaces)
+            removeVatOnlyDir(workspacesRoot);
     };
     const removeSignalCleanup = installSignalCleanup({ onSignal: cleanup });
     // Hoisted so the finally can surface packaging friction even when a throw
@@ -1081,6 +2222,23 @@ export async function runSkillTestHarness(opts) {
     // once merging succeeded; on a pre-merge throw the finally simply finds no
     // friction.json and emits nothing (a no-op).
     let frictionOut;
+    /**
+     * Emit the friction report AT MOST ONCE per run.
+     *
+     * There are now two call sites, and they exist for different reasons that must
+     * not become two reports: the happy path calls it deliberately BEFORE the
+     * baseline delta/banner (so ~1250 rows of friction cannot come between the two —
+     * see the call site), and the `finally` calls it for the throw paths, where no
+     * happy-path call ever happened. A flag rather than a "did we already return normally?"
+     * test, because the throw can land anywhere between the two.
+     */
+    let frictionEmitted = false;
+    const emitFrictionOnce = () => {
+        if (frictionEmitted)
+            return;
+        frictionEmitted = true;
+        emitFrictionReport(frictionOut);
+    };
     try {
         // Step 3: Stage the harness FIRST — the subject's own evals/evals.json lands
         // inside its staged dir, so we must stage before we can locate it. The subject's
@@ -1088,6 +2246,18 @@ export async function runSkillTestHarness(opts) {
         // arrives via opts.subjectSource; build no longer happens in-domain here.
         const resolveCtx = buildResolveCtx(harnessRoot, repoRoot);
         const items = buildStageItems(opts, repoRoot);
+        // Hold a copy of the subject's suite ONLY when this run would have to read it
+        // from there. `resolveEvalSuitePath` prefers the AUTHORED source and falls back
+        // to the hold dir, so when the authored copy exists the held one is never read
+        // — and staging was writing it anyway, putting a second copy of the
+        // `expected_output` answer key in the shared OS temp dir on the common path,
+        // for nothing. The hold dir is 0700 and outside the executor's `--add-dir`
+        // sandbox, but that is Claude's permission model, not an OS boundary: skill
+        // code runs as the same uid and `ls $TMPDIR` finds it. The cheapest answer to
+        // "can the executor reach the answer key" is not to write it there.
+        // Same predicate as `resolveEvalSuitePath`'s rule 1, so the two cannot disagree
+        // about which copy this run uses.
+        const authoredSuiteExists = existsSync(resolveScaffoldEvalsPath(opts, repoRoot, evalsRef));
         const stageResult = await stageHarness({
             harnessRoot,
             items,
@@ -1095,7 +2265,7 @@ export async function runSkillTestHarness(opts) {
             ctx: resolveCtx,
             currentUid,
             evalsSubpath,
-            evalSuiteHoldDir,
+            ...(authoredSuiteExists ? {} : { evalSuiteHoldDir }),
         });
         const { pluginDirs, subjectStagedDir, subjectPluginRoot, skippedOptional } = stageResult;
         emitSkippedOptionalWarning(skippedOptional);
@@ -1106,7 +2276,7 @@ export async function runSkillTestHarness(opts) {
         // staged tree — the answer key must never be reachable by the executor — so this
         // resolves to the AUTHORED source copy, or to the vat-only hold dir when the
         // suite existed nowhere but inside a fetched artifact. Absent from both →
-        // bootstrap a template at the source scaffold and exit 3, so an authored suite
+        // bootstrap a template at the source scaffold and exit 2 with `Reason: bootstrap`, so an authored suite
         // is never overwritten.
         const evalsPath = resolveEvalSuitePath({
             opts,
@@ -1124,24 +2294,29 @@ export async function runSkillTestHarness(opts) {
             const summary = renderPreflightSummary(preflightResult.checks);
             return {
                 harnessPath: harnessRoot,
-                exitCode: SkillTestExitCode.Preflight,
+                exitCode: ExitCode.ERROR,
+                reason: 'preflight',
                 summary: `Preflight failed:\n${summary}`,
             };
         }
         // Step 5.5: Parse the eval suite and stage per-eval input workspaces. The
         // parsed suite is threaded on so the eval loop has the entries without
         // re-reading; declaredEvalCount is derived from it (suite.evals.length).
-        const workspaceStageResult = attemptStageWorkspaces(evalsPath, harnessRoot);
+        const workspaceStageResult = attemptStageWorkspaces(evalsPath, harnessRoot, workspacesRoot, armDirs);
         if ('exitCode' in workspaceStageResult)
             return workspaceStageResult;
-        const { workspacesRoot, declaredEvalCount, suite } = workspaceStageResult;
+        // `workspacesRoot` is resolved above (outside the try, so cleanup can reach it
+        // on a signal); staging returns it unchanged, so only the derived values are
+        // destructured here.
+        const { declaredEvalCount, suite } = workspaceStageResult;
         emitEvalLintWarnings(suite.evals, (opts.declaredExecutables ?? []).map((e) => e.name));
         // Step 6: Enforce the §12 security ack (must pass --i-understand-this-runs-skill-code).
         // Only enforced when not a dry-run and not already acknowledged.
         if (!isAcknowledged(opts)) {
             return {
                 harnessPath: harnessRoot,
-                exitCode: SkillTestExitCode.Preflight,
+                exitCode: ExitCode.ERROR,
+                reason: 'preflight',
                 summary: 'Security acknowledgment required. Pass --i-understand-this-runs-skill-code to proceed.',
             };
         }
@@ -1149,11 +2324,7 @@ export async function runSkillTestHarness(opts) {
         // results/ — grading.json/friction.json/baseline.json come from the merged
         // grader fragments below, never from the (untrusted) model.
         const resultsDir = safePath.joinUnderRoot(harnessRoot, 'results');
-        // 0700, matching the harness root. `grading.json` quotes the executor
-        // transcript verbatim in each expectation's evidence, so whatever the skill
-        // read out of its input files ends up here as text — and unlike the
-        // workspaces, results/ SURVIVES `--keep` by design.
-        mkdirSyncReal(resultsDir, { recursive: true, mode: 0o700 });
+        const isDryRun = opts.dryRun === true;
         // Record what was actually staged & tested (the subject identity + the
         // staged manifest fingerprint + per-entry content hashes) so a run is
         // auditable: which source was resolved, and whether it was rebuilt by run.ts.
@@ -1164,17 +2335,16 @@ export async function runSkillTestHarness(opts) {
             rebuilt: opts.rebuilt === true,
         };
         const provenancePath = safePath.join(resultsDir, 'provenance.json');
-        // eslint-disable-next-line security/detect-non-literal-fs-filename -- our own derived results path
-        writeFileSync(provenancePath, JSON.stringify(provenance, null, 2) + '\n', 'utf-8');
+        openResultsDir({ dryRun: isDryRun, resultsDir, provenancePath, provenance });
         process.stderr.write(`Provenance: ${provenance.fingerprint}\n`);
-        // Resolve the results/ artifact paths and WIPE any prior run's artifacts up
-        // front: harnessRoot is deterministic and may be reused (--keep/--out, or after
-        // a crash before cleanup), so a stale grading/friction/baseline.json must never
-        // leak into — or be echoed by the finally on — this run. `frictionOut` is
-        // hoisted so the finally can echo THIS run's friction (and only after the wipe).
+        // Resolved here, WIPED after the dry-run short-circuit below. harnessRoot is
+        // deterministic and may be reused (--keep/--out, or after a crash before
+        // cleanup), so a stale grading/friction/baseline.json must never leak into — or
+        // be echoed by the finally on — this run; but the wipe is a REAL run's
+        // prerogative, and running it here put it ahead of the dry-run return and of
+        // every remaining pre-pipeline failure point (Step 7.5 token resolution among
+        // them, which is how a FAILING dry run also took the prior artifacts with it).
         const artifacts = resolveArtifactPaths(resultsDir);
-        wipeStaleArtifacts(artifacts);
-        frictionOut = artifacts.frictionOut;
         // Step 7.5: Resolve the executor's declared test env (Features A + B). Token
         // resolution can hard-fail (exit 2) on an unknown ${token}; do it before the
         // dry-run short-circuit so a dry run validates interpolation too. resolvedAuth
@@ -1200,11 +2370,12 @@ export async function runSkillTestHarness(opts) {
         const concurrency = opts.concurrency ?? DEFAULT_CONCURRENCY;
         const modelFlag = buildModelFlag(knobs.model);
         process.stderr.write(`Model: ${knobs.model ?? '(claude default)'} | grader: ${graderModel}\n`);
-        // Step 8: Dry-run short-circuit — return assembled info without spawning.
-        if (opts.dryRun === true) {
+        // Step 8: Dry-run short-circuit — return assembled info without spawning, and
+        // without having touched results/ (see Step 7).
+        if (isDryRun) {
             return {
                 harnessPath: harnessRoot,
-                exitCode: SkillTestExitCode.Ok,
+                exitCode: ExitCode.OK,
                 summary: buildDryRunSummary({
                     wouldBuild: opts.wouldBuild === true,
                     ...(opts.dryRunStagedExistingDist === undefined
@@ -1215,11 +2386,20 @@ export async function runSkillTestHarness(opts) {
                     provenanceEntryCount: provenance.entries.length,
                     modelFlag,
                     evalCount: declaredEvalCount,
+                    // The one fact that doubles the spawn count, and the one the preview was
+                    // blind to — `opts.baseline` was in scope here the whole time.
+                    baseline: opts.baseline === true,
                     concurrency,
                     graderModel,
+                    maxBudgetUsd: knobs.maxBudgetUsd,
                 }),
             };
         }
+        // Past the dry-run return, so this is a REAL run: evict any prior run's
+        // artifacts before writing our own. `frictionOut` is assigned WITH the wipe, so
+        // the `finally` can only ever echo THIS run's friction.
+        wipeStaleArtifacts(artifacts);
+        frictionOut = artifacts.frictionOut;
         // Step 9: Run the vat-owned executor→grader pipeline. ONE secret per-run nonce
         // is stamped into every grader prompt (via stdin) and re-verified per fragment
         // on merge — this is what distinguishes a fragment produced by a grader WE
@@ -1242,8 +2422,15 @@ export async function runSkillTestHarness(opts) {
         const evalCtx = {
             subjectStagedDir,
             workspacesRoot,
+            armDirs,
+            harnessRoot,
             pluginDirs,
             graderOutDir,
+            evalSuiteHoldDir,
+            skillContentNeedles: resolveSkillContentNeedles(subjectStagedDir, suite.evals, {
+                workspacesRoot,
+                armDirs,
+            }),
             runNonce,
             graderModel,
             costAccumulator,
@@ -1261,45 +2448,83 @@ export async function runSkillTestHarness(opts) {
             ...(opts.declaredExecutables === undefined ? {} : { declaredExecutables: opts.declaredExecutables }),
             ...(opts.spawn === undefined ? {} : { spawn: opts.spawn }),
         };
-        // Tier-ordered, cost-tiered fail-fast (issue #145 Phase G): run evals tier by
+        // Tier-ordered, cost-tiered fail-fast: run evals tier by
         // tier (ascending / cheapest first), bounded-parallel WITHIN each tier, and
         // gate BETWEEN tiers — once a cheaper tier fails a gating expectation, the
         // higher (more expensive) tiers are SKIPPED (never graded, never passed), so a
         // broken foundational expectation stops the run from spending on hard tiers.
         // Each tier is one runPipeline: bounded-parallel, retrying a RateLimitSignal
-        // per item. An InternalHarnessError thrown by any executor/grader (timeout/
-        // stall/spawn-error, grader failure, missing fragment, nonce mismatch)
-        // propagates OUT unhandled → mapErrorToExitCode → exit 1; a spawn or grader
-        // break is never laundered into a pass/fail verdict (R1 no-laundering).
-        const { fragments, skipped } = await runEvalsTiered({
+        // per item. An InternalHarnessError thrown by a TREATMENT-arm executor/grader
+        // (timeout/stall/spawn-error, grader failure, missing fragment, nonce mismatch,
+        // declared-count mismatch) propagates OUT unhandled → skillTestFailureReason →
+        // ERROR/internal; a spawn or grader break is never laundered into a pass/fail verdict
+        // (R1 no-laundering). The one grader outcome that IS a verdict — a fragment the
+        // model could not write as valid JSON even after one re-grade — fails that eval
+        // (see unparseableGradingFragment). The same break on the CONTROL arm is
+        // recorded and the run continues — see runEvalWorker for why the two arms are
+        // not symmetric.
+        const { fragments, controlFailures, skipped } = await runEvalsTiered({
             evals: suite.evals,
             baseline: opts.baseline === true,
             runTier: (items) => runPipeline({
                 items,
                 concurrency,
                 worker: (item) => runEvalWorker(item, evalCtx),
+                // The worker rethrows every RateLimitSignal so the pipeline can retry; this
+                // is where an EXHAUSTED one stops being a retry and becomes an outcome. Only
+                // the pipeline knows the budget is spent (see rateLimitExhaustedOutcome).
+                onRetriesExhausted: (item, _index, error) => rateLimitExhaustedOutcome(item, error),
+                ...(opts.rateLimitBackoffMs === undefined ? {} : { onRateLimit: opts.rateLimitBackoffMs }),
             }),
         });
         // Step 10: vat merges the grader fragments and writes grading.json/friction.json/
         // tool-eval.json (and baseline.json for a baseline run), then reconciles the
         // authoritative prose-expectation verdict from the WITH-arm per-expectation
         // `passed` flags — NOT any self-reported summary. An executor CLEAN failure
-        // reaches here as a FAILing fragment → composite verdict → exit 4.
-        const { verdict, toolEval } = writeRunArtifactsAndReconcile(fragments, runNonce, artifacts);
+        // reaches here as a FAILing fragment → composite verdict → exit FINDINGS.
+        const { verdict, toolEval, baselineReport } = writeRunArtifactsAndReconcile({
+            fragments,
+            runNonce,
+            paths: artifacts,
+            signals: contaminationSignalsFor(evalCtx),
+            controlFailures,
+            // `skipped` was in scope here and simply not passed, which is how the delta
+            // block and its printed line came to describe a truncated run as a complete one.
+            skipped,
+            subjectInPlace: opts.subjectInPlace,
+        });
         // D2 fail-closed gate: vat is the SOLE writer of results/, so a missing/unparseable/
-        // invalid grading.json, friction.json, or tool-eval.json after the merge is a HARNESS
-        // bug (exit 1), never a skill fault. A SEPARATE explicit check, not a prompt invariant.
-        assertVatWroteArtifacts(artifacts);
+        // invalid grading.json, friction.json, tool-eval.json — or, on a baseline run,
+        // baseline.json — after the merge is a HARNESS bug (exit 1), never a skill fault.
+        // A SEPARATE explicit check, not a prompt invariant.
+        assertVatWroteArtifacts(artifacts, opts.baseline === true);
+        // Friction FIRST, then the baseline report — the ORDER is the point. Friction is
+        // capped at 50 lines of 2000 characters, so emitting it after the ⚠️ contaminated
+        // banner (which is what the `finally` used to do) could put ~1250 terminal rows
+        // between the operator and the only line saying the delta above is meaningless.
+        // Emitting it here leaves nothing of vat's own between the delta and its caveat;
+        // `emitFrictionOnce` keeps the `finally` call a no-op so a normal run reports it
+        // exactly once. The CLI still appends `Harness:`/`Results:`/`Workspaces:`/
+        // `Summary:` after the harness returns, so the banner is last-but-2-to-4, not
+        // last — a bounded tail, unlike the friction report.
+        emitFrictionOnce();
+        emitBaselineReport(baselineReport);
         // Composite verdict: the run passes only when BOTH output expectations AND every
         // tool-expectation verdict passed — so an output all-pass where a `mustRun` never
-        // ran yields FAIL → exit 4. Tool verdicts stay in tool-eval.json (C2); the composite
+        // ran yields FAIL → exit FINDINGS. Tool verdicts stay in tool-eval.json (C2); the composite
         // combines the two channels only here, at the exit-code layer. A fail-fast run
         // that SKIPPED higher tiers is never a pass: skipped ≠ passed forces allPassed
-        // false → exit 4 (never downgraded to 0 by the composite path).
+        // false → exit FINDINGS (never downgraded to 0 by the composite path).
         const allPassed = resolveCompositeAllPassed(verdict.allPassed, toolEval, skipped);
         const summary = buildRunSummaryWithSkips(verdict, toolEval, allPassed, skipped, costAccumulator);
         return {
             harnessPath: harnessRoot,
+            // Reported ONLY when the workspaces outlive the run. The `finally` below
+            // runs cleanup BEFORE this value ever reaches a caller, so reporting it
+            // unconditionally handed the operator a path to a directory that no longer
+            // existed. Same predicate as cleanup's, by construction.
+            ...(retainWorkspaces ? { workspacesPath: workspacesRoot } : {}),
+            resultsPath: resultsDir,
             exitCode: verdictExitCode(allPassed, opts.tolerateEvalFailure === true),
             summary,
         };
@@ -1308,8 +2533,10 @@ export async function runSkillTestHarness(opts) {
         // Surface any packaging-fidelity friction VAT merged into friction.json to
         // STDERR BEFORE cleanup removes the harness dir. Best-effort + stderr-only
         // (stdout stays machine-readable); a no-op when the path is unassigned (a
-        // pre-merge throw) or friction.json is absent or empty.
-        emitFrictionReport(frictionOut);
+        // pre-merge throw) or friction.json is absent or empty. On a run that reached
+        // the verdict this is ALREADY a no-op — the happy path emits it earlier, ahead
+        // of the baseline banner — so what is left here is the throw paths.
+        emitFrictionOnce();
         // Remove the signal handlers first (no listener leak across runs), then run
         // the same cleanup: release the lock, remove the harness dir, and remove the
         // vat-only grader dir (which lives outside harnessRoot).
@@ -1317,4 +2544,62 @@ export async function runSkillTestHarness(opts) {
         cleanup();
     }
 }
+/**
+ * The test-facing seam — every helper the unit tests reach that is NOT part of
+ * this module's API.
+ *
+ * These used to be forty-four separate `export`s, indistinguishable at the export
+ * keyword from the seven names the skill-test barrel actually publishes, so
+ * "is this API?" was a question for a JSDoc rather than a grep. Now it is a
+ * grep: a name here is reachable by a test and by nothing else. Nothing under
+ * `src/` imports `__internal` and no barrel re-exports it — a dev-tools test
+ * holds both lines. Types stay exported on their own: an object cannot carry
+ * them, and a type is not a call site.
+ */
+export const __internal = {
+    assertVatWroteArtifacts,
+    buildContaminationInput,
+    buildContaminationSignalsInput,
+    buildEvalWorkItems,
+    buildFlagParseProbe,
+    buildPreflightInput,
+    buildResolveCtx,
+    buildRunSummary,
+    buildRunSummaryWithSkips,
+    computeCompositeVerdict,
+    detectItemPluginLayout,
+    emitFrictionReport,
+    FLAG_PROBE_SENTINEL,
+    formatBaselineReport,
+    formatRunCostSuffix,
+    gradedCounts,
+    helpTextDeclaresFlag,
+    makeStageItem,
+    mintArmWorkspaceDirs,
+    partitionFragmentsByArm,
+    recordSessionCost,
+    rejectedArtifactPath,
+    removeVatOnlyDir,
+    renderPreflightSummary,
+    resolveArtifactPaths,
+    resolveCompositeAllPassed,
+    resolveHarnessLocation,
+    resolveKnobs,
+    resolvePerEvalWorkspaceDir,
+    resolveScaffoldEvalsPath,
+    resolveSkillContentNeedles,
+    resolveStallMs,
+    resolveTimeoutMs,
+    resolveWorkspacesRoot,
+    RETAINED_RESULTS_DIRNAME,
+    stageWorkspacesForRun,
+    subjectSkillName,
+    swallowCleanupFailure,
+    wipeStaleArtifacts,
+    withoutGraderContamination,
+    cleanupHarness,
+    buildStageItems,
+    resolveGraderOutDir,
+    resolveEvalSuitePath,
+};
 //# sourceMappingURL=run-harness.js.map