@herbertgao/pi-extensions 2026.10.1 → 2026.10.3

package/node_modules/pi-lens/dist/clients/dispatch/dispatcher.js

@@ -1,1269 +0,0 @@
-/**
- * Declarative Tool Dispatcher for pi-lens
- *
- * Redesigned to handle the full complexity of pi-lens's tool_result handler:
- * - Multiple tools with different semantics (blocking, warning, silent)
- * - Delta mode (baseline tracking)
- * - Autofix handling
- * - Output aggregation and formatting
- *
- * Key abstractions:
- * - RunnerDefinition: A tool that can be run
- * - Diagnostic: Structured issue representation
- * - OutputSemantic: How to display (blocking, warning, silent, etc.)
- * - BaselineStore: Track pre-existing issues for delta mode
- */
-import { logExtension } from "../extension-log.js";
-import * as fs from "node:fs";
-import * as path from "node:path";
-import { recordRunner } from "../widget-state.js";
-import { incrementDegradationCount, recordDegradationOnce, } from "../degradation-ledger.js";
-import { detectFileKind } from "../file-kinds.js";
-import { detectFileRole } from "../file-role.js";
-import { classifyGeneratedOrArtifactDetailed, } from "../generated-artifacts.js";
-import { isTestFile } from "../file-utils.js";
-import { getPrimaryDispatchGroup } from "../language-policy.js";
-import { resolveLanguageRootForFile } from "../language-profile.js";
-import { logLatency, phaseFinished, phaseStarted } from "../latency-logger.js";
-import { isSpawnableCommand } from "../installer/index.js";
-import { isAbsoluteFilePath, isWindowsPath, normalizeEphemeralMapKey, normalizeMapKey, } from "../path-utils.js";
-import { loadPiLensProjectConfig } from "../project-lens-config.js";
-import { RUNTIME_CONFIG, getRunnerTimeoutFloorMs } from "../runtime-config.js";
-import { probeToolAsync } from "../tool-probe.js";
-import { classifyDiagnostic } from "./diagnostic-taxonomy.js";
-import { classifyProbeFailure, logAvailabilityDecision, startHostStallSampler, transientRetryDelayMs, } from "./runners/utils/availability-policy.js";
-import { recordAvailabilityProbeOverrun, getDispatchAvailabilityGeneration, } from "./runners/utils/runner-helpers.js";
-import { createAvailabilityProbeFlight } from "../availability-probe-flight.js";
-import { applyDispositions } from "../diagnostic-dispositions.js";
-import { applyInlineSuppressions } from "./inline-suppressions.js";
-import { getToolPlan } from "./plan.js";
-import { resolveRunnerPath } from "./runner-context.js";
-import { classifyObservedRunner, COLLECT_LATER_THRESHOLD_MS, observeRunnerLatency, } from "./collect-later-tier.js";
-import { deferRunnerFindings } from "./pending-runner-findings.js";
-import { applyRulePolicy, rulePolicyMapFromConfig } from "./rule-policy.js";
-import { getToolProfile } from "./tool-profile.js";
-import { isRunnerSkipReason } from "./types.js";
-const dispatcherProbeFlights = createAvailabilityProbeFlight({ generation: () => getDispatchAvailabilityGeneration() });
-import { formatDiagnostics } from "./utils/format-utils.js";
-// --- Runner Registry ---
-export class RunnerRegistry {
-    runners = new Map();
-    register(runner) {
-        if (this.runners.has(runner.id))
-            return;
-        this.runners.set(runner.id, runner);
-    }
-    get(id) {
-        return this.runners.get(id);
-    }
-    getForKind(kind, filePath) {
-        const matching = [];
-        const isTest = filePath ? isTestFile(filePath) : false;
-        for (const runner of this.runners.values()) {
-            if (isTest && runner.skipTestFiles)
-                continue;
-            if (runnerAppliesToKind(runner, kind)) {
-                matching.push(runner);
-            }
-        }
-        return matching.sort((a, b) => a.priority - b.priority);
-    }
-    list() {
-        return Array.from(this.runners.values());
-    }
-    clear() {
-        this.runners.clear();
-    }
-}
-function runnerAppliesToKind(runner, kind) {
-    return (runner.appliesTo.length === 0 ||
-        (kind !== undefined && runner.appliesTo.includes(kind)));
-}
-// --- Tool Availability Cache ---
-/**
- * Normalize a command name to a FactStore session key.
- * Strips .cmd/.exe suffixes (case-insensitive) and lowercases,
- * then prefixes with "session.toolCache.".
- */
-export function normalizeCacheKey(cmd) {
-    const normalized = cmd.replace(/\.(cmd|exe)$/i, "").toLowerCase();
-    return `session.toolCache.${normalized}`;
-}
-/** Probe budget for the generic `hasTool` availability check, ms. */
-const TOOL_PROBE_TIMEOUT_MS = 5000;
-/**
- * Session fact holding the epoch ms after which a TRANSIENT verdict for a
- * command may be re-probed. Kept beside the boolean fact so both share the
- * session's lifetime — no second global to reset (#1476).
- */
-function transientRetryKey(command) {
-    return `${normalizeCacheKey(command)}.retryAt`;
-}
-/**
- * Consecutive transient verdicts for a command, so the cooldown ESCALATES the
- * way the policy documents (30 s, 60 s, 120 s … capped at 5 min) instead of
- * sitting flat at 30 s.
- *
- * This is the highest-traffic availability consumer in the product. A flat
- * cooldown here means a permanently sick host is re-probed every 30 s per
- * command for the whole session — ten times the storm the policy claims to
- * bound, and the one place the bound matters most. `createAvailabilityChecker`
- * tracks the same counter; this keeps the two seams telling one story.
- */
-function transientAttemptsKey(command) {
-    return `${normalizeCacheKey(command)}.transientAttempts`;
-}
-/**
- * Is `command` usable right now? Cached per session.
- *
- * Latch policy (#1467/#1476): a `false` from a genuine absence is durable and
- * is remembered for the session. A `false` from a TIMED-OUT probe is not — this
- * is the highest-traffic availability consumer in the codebase, and caching a
- * host stall here silently disabled a healthy tool for every later dispatch in
- * the session. A transient verdict instead holds only for a bounded cooldown,
- * which also stops a sick host from being re-probed on every dispatch.
- */
-export async function checkToolAvailability(command, facts) {
-    const key = normalizeCacheKey(command);
-    const cached = facts.getSessionFact(key);
-    if (cached !== undefined) {
-        return cached;
-    }
-    const retryAt = facts.getSessionFact(transientRetryKey(command));
-    if (retryAt !== undefined && Date.now() < retryAt)
-        return false;
-    // A command that isn't even on disk can't pass a --version probe; the ~μs
-    // stat/PATH walk saves a guaranteed-to-fail spawn round-trip per cold tool.
-    if (!(await isSpawnableCommand(command))) {
-        facts.setSessionFact(key, false);
-        return false;
-    }
-    try {
-        // The budget is enforced by a HOST-side timer, so measure the loop stall
-        // that overlapped the window and let the shared classifier read it.
-        const sampler = startHostStallSampler();
-        const startedAt = Date.now();
-        let result;
-        let hostStallMs;
-        let probeJoined = false;
-        try {
-            const shared = dispatcherProbeFlights.run(`dispatcher:${key}`, () => probeToolAsync(command, ["--version"], {
-                timeout: TOOL_PROBE_TIMEOUT_MS,
-            }));
-            probeJoined = shared.joined;
-            result = await shared.promise;
-        }
-        finally {
-            hostStallMs = sampler.stop();
-        }
-        const elapsedMs = Date.now() - startedAt;
-        recordAvailabilityProbeOverrun(command, key, elapsedMs, TOOL_PROBE_TIMEOUT_MS, result.failure);
-        if (result.status === 0) {
-            facts.setSessionFact(key, true);
-            // The tool answered: retire the cooldown facts rather than leaving a
-            // stale retry deadline behind a `true`.
-            facts.setSessionFact(transientRetryKey(command), 0);
-            facts.setSessionFact(transientAttemptsKey(command), 0);
-            logAvailabilityDecision({
-                tool: command,
-                verdict: "available",
-                outcome: "success",
-                cause: "ok",
-                elapsedMs,
-                latched: true,
-                classifiedBy: probeJoined ? "joined" : "probe",
-                hostStallMs,
-                budgetMs: TOOL_PROBE_TIMEOUT_MS,
-            });
-            return true;
-        }
-        // `missing` for anything unclassified preserves the pre-#1476 meaning of
-        // a non-zero exit: durable, cached for the session.
-        const { outcome, cause, evidence } = classifyProbeFailure(result, {
-            hostStallMs,
-            unclassifiedFailureOutcome: "missing",
-        });
-        let retryAfterMs;
-        if (outcome === "transient") {
-            const attempts = (facts.getSessionFact(transientAttemptsKey(command)) ?? 0) + 1;
-            facts.setSessionFact(transientAttemptsKey(command), attempts);
-            retryAfterMs = transientRetryDelayMs(attempts, cause);
-            facts.setSessionFact(transientRetryKey(command), Date.now() + retryAfterMs);
-        }
-        else {
-            facts.setSessionFact(key, false);
-            facts.setSessionFact(transientAttemptsKey(command), 0);
-        }
-        logAvailabilityDecision({
-            tool: command,
-            verdict: "unavailable",
-            outcome,
-            cause,
-            elapsedMs,
-            latched: outcome !== "transient",
-            hostStallMs,
-            ...(retryAfterMs !== undefined && { retryAfterMs }),
-            budgetMs: TOOL_PROBE_TIMEOUT_MS,
-            classifiedBy: probeJoined ? "joined" : "probe",
-            evidence,
-        });
-        return false;
-    }
-    catch {
-        facts.setSessionFact(key, false);
-        return false;
-    }
-}
-// --- Dispatch Context Factory ---
-function readFilePrefix(filePath, maxBytes = 4096) {
-    let fd;
-    try {
-        fd = fs.openSync(filePath, "r");
-        const buffer = Buffer.alloc(maxBytes);
-        const bytesRead = fs.readSync(fd, buffer, 0, maxBytes, 0);
-        return buffer.subarray(0, bytesRead).toString("utf8");
-    }
-    catch {
-        return undefined;
-    }
-    finally {
-        if (fd !== undefined) {
-            try {
-                fs.closeSync(fd);
-            }
-            catch {
-                // ignore close errors
-            }
-        }
-    }
-}
-export function createDispatchContext(filePath, cwd, pi, facts, blockingOnly, modifiedRanges, 
-/** Authoritative workspace root; `cwd` may be a nested language root. */
-projectRoot, 
-/** Ordered per-file pipeline token, when this is a post-write dispatch. */
-writeIndex, 
-/** Runtime telemetry identity, when known (#1448) — threaded to the
- * worklog append; see DispatchContext.telemetryModel's doc. */
-telemetryModel, telemetryProvider) {
-    const absoluteFilePath = resolveRunnerPath(cwd, filePath);
-    const normalizedProjectRoot = normalizeMapKey(path.resolve(projectRoot ?? cwd));
-    const normalizedCwd = normalizeMapKey(resolveLanguageRootForFile(absoluteFilePath, cwd));
-    const normalizedFilePath = normalizeMapKey(absoluteFilePath);
-    // `resolveRunnerPath` intentionally normalizes map keys, which folds a
-    // literal POSIX backslash into a separator. Preserve that one proven
-    // classifier distinction while still resolving relative inputs from `cwd`.
-    const kindFilePath = !isWindowsPath(filePath) && filePath.includes("\\")
-        ? path.resolve(cwd, filePath)
-        : absoluteFilePath;
-    const kind = detectFileKind(kindFilePath);
-    const contentPrefix = readFilePrefix(normalizedFilePath);
-    const fileRole = detectFileRole(normalizedFilePath, contentPrefix);
-    // Captured once here so the generated short-circuit below can emit a
-    // `dispatch_skipped_generated` record carrying the deciding evidence tier
-    // and the measured line-shape statistic — without re-reading the file
-    // (refs #2346). The content passed is the same 4096-byte prefix already
-    // read for role detection, so this classification costs no extra I/O.
-    const generatedDetail = fileRole === "generated"
-        ? classifyGeneratedOrArtifactDetailed(normalizedFilePath, {
-            content: contentPrefix,
-            includeDeclarations: false,
-        })
-        : undefined;
-    const projectConfig = loadPiLensProjectConfig(normalizedCwd);
-    return {
-        filePath: normalizedFilePath,
-        projectRoot: normalizedProjectRoot,
-        cwd: normalizedCwd,
-        kind,
-        fileRole,
-        generatedEvidence: generatedDetail?.evidence,
-        generatedLineShapeMean: generatedDetail?.lineShapeMean,
-        pi,
-        autofix: false,
-        deltaMode: !pi.getFlag("no-delta"),
-        facts,
-        projectConfig,
-        blockingOnly,
-        modifiedRanges,
-        writeIndex,
-        telemetryModel,
-        telemetryProvider,
-        toolCwdMemo: {},
-        async hasTool(command) {
-            return checkToolAvailability(command, facts);
-        },
-        log(message, level) {
-            // #1333: pi owns the terminal — a runner advisory must never be a raw
-            // write. Every DispatchContext.log line lands in extension.log instead.
-            logExtension({
-                subsystem: "dispatch",
-                // exactOptionalPropertyTypes: an omitted level must stay omitted.
-                ...(level !== undefined && { level }),
-                message,
-                metadata: { filePath: normalizedFilePath, kind },
-            });
-        },
-    };
-}
-// --- Delta Mode Logic ---
-/**
- * Filter diagnostics to only show NEW issues (delta mode)
- */
-function filterDelta(after, before, keyFn) {
-    const beforeSet = new Set((before ?? []).map(keyFn));
-    const afterSet = new Set(after.map(keyFn));
-    const fixed = (before ?? []).filter((d) => !afterSet.has(keyFn(d)));
-    const newItems = after.filter((d) => !beforeSet.has(keyFn(d)));
-    return { new: newItems, fixed };
-}
-function semanticRank(semantic) {
-    if (semantic === "blocking")
-        return 4;
-    if (semantic === "warning")
-        return 3;
-    if (semantic === "fixed")
-        return 2;
-    if (semantic === "silent")
-        return 1;
-    return 0;
-}
-function toolPriority(tool, defectClass) {
-    return getToolProfile(tool, defectClass).dedupPriority;
-}
-function dedupeOverlappingDiagnostics(diagnostics) {
-    const byKey = new Map();
-    for (const d of diagnostics) {
-        const defectClass = d.defectClass ?? classifyDiagnostic(d);
-        const line = d.line ?? 1;
-        const column = d.column ?? 1;
-        const ruleKey = d.rule || d.id || "unknown";
-        const key = `${d.filePath}:${line}:${column}:${defectClass}:${ruleKey}`;
-        const current = byKey.get(key);
-        if (!current) {
-            byKey.set(key, { ...d, defectClass });
-            continue;
-        }
-        const currScore = semanticRank(current.semantic) * 100 +
-            toolPriority(current.tool, defectClass);
-        const nextScore = semanticRank(d.semantic) * 100 + toolPriority(d.tool, defectClass);
-        if (nextScore > currScore) {
-            byKey.set(key, { ...d, defectClass });
-        }
-    }
-    return [...byKey.values()];
-}
-function suppressLintOverlapsWithLsp(diagnostics) {
-    const lspBySpanClass = new Set();
-    const lspByLine = new Set();
-    const isLintTool = (tool) => {
-        return getToolProfile(tool).lintLike;
-    };
-    for (const d of diagnostics) {
-        if (d.tool !== "lsp")
-            continue;
-        const line = d.line ?? 1;
-        const defectClass = d.defectClass ?? classifyDiagnostic(d);
-        lspBySpanClass.add(`${d.filePath}:${line}:${defectClass}`);
-        lspByLine.add(`${d.filePath}:${line}`);
-    }
-    if (lspByLine.size === 0)
-        return diagnostics;
-    return diagnostics.filter((d) => {
-        if (d.tool === "lsp")
-            return true;
-        if (!isLintTool(d.tool))
-            return true;
-        if (d.semantic === "blocking" || d.severity === "error")
-            return true;
-        const line = d.line ?? 1;
-        const defectClass = d.defectClass ?? classifyDiagnostic(d);
-        const key = `${d.filePath}:${line}:${defectClass}`;
-        if (lspBySpanClass.has(key))
-            return false;
-        // Conservative fallback for unclassified overlap at same line.
-        if (defectClass === "unknown") {
-            return !lspByLine.has(`${d.filePath}:${line}`);
-        }
-        return true;
-    });
-}
-/**
- * Dockerfile overlap dedup (#131 Mode 2): hadolint and `trivy config` both flag
- * a few of the same Dockerfile issues (e.g. `:latest`, running as root) with
- * different rule ids, which the rule-keyed `dedupeOverlappingDiagnostics` can't
- * collapse. Keep hadolint authoritative on the lines it covers and drop the
- * trivy-config finding there — trivy still contributes the security checks
- * hadolint lacks (on other lines), and all Kubernetes findings (no hadolint
- * diagnostics exist for YAML, so none are suppressed). Exported for unit tests.
- */
-export function suppressTrivyConfigDockerOverlap(diagnostics) {
-    const hadolintLines = new Set();
-    for (const d of diagnostics) {
-        if (d.tool === "hadolint") {
-            hadolintLines.add(`${d.filePath}:${d.line ?? 1}`);
-        }
-    }
-    if (hadolintLines.size === 0)
-        return diagnostics;
-    return diagnostics.filter((d) => d.tool !== "trivy-config" ||
-        !hadolintLines.has(`${d.filePath}:${d.line ?? 1}`));
-}
-function isUnusedValueDiagnostic(d) {
-    const raw = `${d.id ?? ""} ${d.rule ?? ""} ${d.message ?? ""}`.toLowerCase();
-    if (raw.includes("no-unused"))
-        return true;
-    if (/\b(6133|6192|6196)\b/.test(raw))
-        return true;
-    const rule = String(d.rule ?? "").toLowerCase();
-    if (rule.includes("unused"))
-        return true;
-    const message = d.message.toLowerCase();
-    return (message.includes("is declared but its value is never read") ||
-        message.includes("is assigned a value but never used") ||
-        message.includes("declared but never used") ||
-        message.includes("unused"));
-}
-function promoteDeltaUnusedToBlockers(diagnostics) {
-    return diagnostics.map((d) => {
-        if (!isUnusedValueDiagnostic(d))
-            return d;
-        if (d.semantic === "blocking" || d.severity === "error")
-            return d;
-        return {
-            ...d,
-            severity: "error",
-            semantic: "blocking",
-            fixSuggestion: d.fixSuggestion ??
-                "Remove the unused declaration or rename with '_' prefix if intentionally unused.",
-        };
-    });
-}
-function buildCoverageNotice(ctx, runnerLatencies) {
-    if (!ctx.kind)
-        return undefined;
-    const lspEnabled = !ctx.pi.getFlag("no-lsp");
-    const primary = getPrimaryDispatchGroup(ctx.kind, lspEnabled);
-    if (!primary || primary.runnerIds.length === 0)
-        return undefined;
-    const relevant = runnerLatencies.filter((r) => primary.runnerIds.includes(r.runnerId));
-    if (relevant.length === 0)
-        return undefined;
-    // #1867 catalog shape 4: this is correlation, not classification. The LSP
-    // touch already decided which scanner publications were absent for these
-    // bytes; the runner and latency assembly only preserve that exact set.
-    const unconfirmedServerIds = [
-        ...new Set(relevant.flatMap((r) => r.unconfirmedServerIds ?? [])),
-    ];
-    if (unconfirmedServerIds.length > 0) {
-        const deferredIds = new Set(relevant.flatMap((r) => r.deferredServerIds ?? []));
-        const markerFor = (ids) => {
-            const shown = ids.slice(0, 4);
-            const remainder = ids.length - shown.length;
-            return `${shown.join(", ")}${remainder > 0 ? ` +${remainder}` : ""}`;
-        };
-        const deferredServerIds = unconfirmedServerIds.filter((id) => deferredIds.has(id));
-        const silentServerIds = unconfirmedServerIds.filter((id) => !deferredIds.has(id));
-        // The marker describes this exact scanner set. A scanner can recover
-        // while another goes dark on the same file, so the set belongs in the
-        // session dedupe identity rather than only kind and path.
-        // #2016: these are SCANNER IDS, not filesystem paths. `normalizeMapKey`
-        // would realpath each one; on Windows that fails, falls through to
-        // `resolveNonExisting`, and resolves the id against the CURRENT process
-        // cwd, so the dedupe key differed by platform and by cwd (the #2219
-        // non-path-sentinel class). The cheap syntactic fold is what this
-        // session-scoped dedupe key actually needs. Code-unit comparator: the
-        // sorted set is a KEY, so ordering must be deterministic across locales
-        // — localeCompare is not.
-        const dedupeSet = (ids) => [...new Set(ids)]
-            .map(normalizeEphemeralMapKey)
-            .sort((a, b) => Number(a > b) - Number(a < b))
-            .join(",");
-        // #2810 round 4: the PARTITION is part of the identity, not just the
-        // scanner set. The same scanner can be silent on one edit and marked for
-        // late delivery on the next; keying on the unconfirmed set alone showed
-        // the session whichever message came first and suppressed the other, so a
-        // scanner that recovered a delivery path (or lost one) kept the stale
-        // wording for the rest of the session.
-        const onceKey = `${ctx.kind}:${ctx.filePath}:${dedupeSet(silentServerIds)}|${dedupeSet(deferredServerIds)}`;
-        if (coverageNoticeSeen.has(onceKey))
-            return undefined;
-        coverageNoticeSeen.add(onceKey);
-        const coverageParts = [];
-        if (deferredServerIds.length > 0) {
-            coverageParts.push(`coverage: ${markerFor(deferredServerIds)} deferred — diagnostics are incomplete; findings arrive at turn end if the scan lands.`);
-        }
-        if (silentServerIds.length > 0) {
-            coverageParts.push(`coverage: ${markerFor(silentServerIds)} silent — diagnostics are incomplete (not a clean result).`);
-        }
-        return {
-            id: `coverage-partial:${ctx.kind}:${path.basename(ctx.filePath)}`,
-            message: coverageParts.join("\n"),
-            filePath: ctx.filePath,
-            severity: "warning",
-            semantic: "warning",
-            tool: "pi-lens",
-        };
-    }
-    // Check primary runners first
-    const primaryHasCoverage = relevant.some((r) => r.status === "succeeded" || r.status === "failed");
-    if (primaryHasCoverage)
-        return undefined;
-    const allPrimarySkipped = relevant.every((r) => r.status === "skipped" ||
-        r.status === "when_skipped" ||
-        r.status === "test_file_skipped");
-    if (!allPrimarySkipped)
-        return undefined;
-    const plan = getToolPlan(ctx.kind);
-    const fallbackRunnerIds = new Set((plan?.groups ?? [])
-        .filter((group) => !group.runnerIds.every((runnerId) => primary.runnerIds.includes(runnerId)))
-        .flatMap((group) => group.runnerIds)
-        .filter((runnerId) => !primary.runnerIds.includes(runnerId)));
-    // Structural-only runners (tree-sitter, ast-grep) are not substitutes
-    // for real linters — don't suppress the notice if only they ran.
-    const STRUCTURAL_RUNNERS = new Set([
-        "tree-sitter",
-        "ast-grep-napi",
-        "spellcheck",
-        "fact-rules",
-        "opengrep",
-    ]);
-    const anyLinterHasCoverage = runnerLatencies.some((r) => fallbackRunnerIds.has(r.runnerId) &&
-        !STRUCTURAL_RUNNERS.has(r.runnerId) &&
-        (r.status === "succeeded" || r.status === "failed"));
-    if (anyLinterHasCoverage)
-        return undefined;
-    const onceKey = `${ctx.kind}:${ctx.filePath}`;
-    if (coverageNoticeSeen.has(onceKey))
-        return undefined;
-    coverageNoticeSeen.add(onceKey);
-    return {
-        id: `coverage-unavailable:${ctx.kind}:${path.basename(ctx.filePath)}`,
-        message: `Pi-lens ${ctx.kind} analysis unavailable — language tools are missing or the LSP server isn't ready yet, so this file was not fully checked (not a clean result).`,
-        filePath: ctx.filePath,
-        severity: "warning",
-        semantic: "warning",
-        tool: "pi-lens",
-    };
-}
-const latencyReports = [];
-const coverageNoticeSeen = new Set();
-// One `dispatch_skipped_generated` phase record per file per process (refs
-// #2346): a generated file dispatched repeatedly must not spam latency.log,
-// so only the first skip of each file emits the row; the degradation ledger
-// below still tallies every repetition with a bounded per-subject count.
-const generatedSkipRecorded = new Set();
-export function getLatencyReports() {
-    return [...latencyReports];
-}
-export function clearLatencyReports() {
-    latencyReports.length = 0;
-}
-export function clearCoverageNoticeState() {
-    coverageNoticeSeen.clear();
-    generatedSkipRecorded.clear();
-}
-export function formatLatencyReport(report) {
-    const lines = [];
-    lines.push(`\n═══════════════════════════════════════════════════════════════`);
-    lines.push(`📊 DISPATCH LATENCY REPORT: ${report.filePath.split("/").pop()}`);
-    lines.push(`   Kind: ${report.fileKind || "unknown"} | Total: ${report.totalDurationMs}ms`);
-    lines.push(`───────────────────────────────────────────────────────────────`);
-    lines.push(`Runner                          Duration  Status    Issues  Semantic`);
-    lines.push(`───────────────────────────────────────────────────────────────`);
-    for (const r of report.runners) {
-        const name = r.runnerId.padEnd(30);
-        const dur = `${r.durationMs}ms`.padStart(8);
-        const status = r.status.padStart(9);
-        const issues = String(r.diagnosticCount).padStart(6);
-        const sem = r.semantic.padStart(8);
-        const slowMarker = r.durationMs > 500 ? " 🔥" : r.durationMs > 100 ? " ⚡" : "";
-        lines.push(`${name}${dur}${status}${issues}${sem}${slowMarker}`);
-    }
-    lines.push(`───────────────────────────────────────────────────────────────`);
-    lines.push(`Total: ${report.runners.length} runners | Stopped early: ${report.stoppedEarly}`);
-    lines.push(`Diagnostics: ${report.totalDiagnostics} (${report.blockers} blockers, ${report.warnings} warnings)`);
-    // Show top 3 slowest
-    const sorted = [...report.runners].sort((a, b) => b.durationMs - a.durationMs);
-    if (sorted.length > 0 && sorted[0].durationMs > 100) {
-        lines.push(`\n🐌 Slowest runners:`);
-        for (const r of sorted.slice(0, 3)) {
-            if (r.durationMs > 50) {
-                lines.push(`   ${r.runnerId}: ${r.durationMs}ms (${r.status})`);
-            }
-        }
-    }
-    lines.push(`═══════════════════════════════════════════════════════════════`);
-    return lines.join("\n");
-}
-/**
- * Execute all runners in a single group.
- *
- * - mode "fallback": run runners sequentially and stop at the first
- *   one that succeeds (returns status !== "skipped").
- * - mode "all" (default): run all runners in the group sequentially
- *   and collect every diagnostic.
- *
- * Groups themselves are run in parallel by dispatchForFile, so this
- * function must NOT mutate shared state.
- */
-async function runGroup(ctx, group, registry, onRunnerResult) {
-    const diagnostics = [];
-    const latencies = [];
-    let hadBlocker = false;
-    // Filter runners by kind if specified
-    const runnerIds = group.filterKinds
-        ? group.runnerIds.filter((id) => {
-            const runner = registry.get(id);
-            return runner && ctx.kind && group.filterKinds?.includes(ctx.kind);
-        })
-        : group.runnerIds;
-    const semantic = group.semantic ?? "warning";
-    for (const runnerId of runnerIds) {
-        const runnerStart = Date.now();
-        const runner = registry.get(runnerId);
-        if (!runner) {
-            latencies.push({
-                runnerId,
-                startTime: runnerStart,
-                endTime: Date.now(),
-                durationMs: 0,
-                status: "skipped",
-                diagnosticCount: 0,
-                semantic: "unknown",
-            });
-            logLatency({
-                type: "runner",
-                filePath: ctx.filePath,
-                runnerId,
-                durationMs: 0,
-                status: "not_registered",
-                diagnosticCount: 0,
-                semantic: "unknown",
-            });
-            continue;
-        }
-        // Same skipTestFiles gate RunnerRegistry.getForKind applies for the
-        // per-edit path (#2337): the plan/group path resolves runners by id via
-        // registry.get() instead, which bypasses getForKind entirely, so a
-        // runner declaring skipTestFiles never had it enforced here.
-        if (runner.skipTestFiles && isTestFile(ctx.filePath)) {
-            latencies.push({
-                runnerId,
-                startTime: runnerStart,
-                endTime: Date.now(),
-                durationMs: Date.now() - runnerStart,
-                status: "test_file_skipped",
-                diagnosticCount: 0,
-                semantic: "none",
-            });
-            logLatency({
-                type: "runner",
-                filePath: ctx.filePath,
-                runnerId,
-                durationMs: 0,
-                status: "test_file_skipped",
-                diagnosticCount: 0,
-                semantic: "none",
-            });
-            continue;
-        }
-        // Keep explicit groups aligned with RunnerRegistry.getForKind(): an empty
-        // appliesTo list means every kind, while a populated list must contain the
-        // current kind. A filtered runner remains visible as skipped telemetry so
-        // language mismatches cannot be mistaken for runner failures.
-        const appliesToCurrentKind = runnerAppliesToKind(runner, ctx.kind);
-        if (!appliesToCurrentKind) {
-            const runnerEnd = Date.now();
-            latencies.push({
-                runnerId,
-                startTime: runnerStart,
-                endTime: runnerEnd,
-                durationMs: 0,
-                status: "skipped",
-                diagnosticCount: 0,
-                semantic: "unknown",
-            });
-            logLatency({
-                type: "runner",
-                filePath: ctx.filePath,
-                runnerId,
-                durationMs: 0,
-                status: "skipped",
-                diagnosticCount: 0,
-                semantic: "unknown",
-                metadata: { reason: "applies_to", kind: ctx.kind },
-            });
-            continue;
-        }
-        // Check preconditions
-        let shouldRun = true;
-        if (runner.when) {
-            try {
-                shouldRun = await runner.when(ctx);
-            }
-            catch (error) {
-                ctx.log(`Runner ${runner.id} precondition failed: ${error}`);
-                shouldRun = false;
-            }
-        }
-        if (!shouldRun) {
-            latencies.push({
-                runnerId,
-                startTime: runnerStart,
-                endTime: Date.now(),
-                durationMs: Date.now() - runnerStart,
-                status: "when_skipped",
-                diagnosticCount: 0,
-                semantic: runner.id,
-            });
-            logLatency({
-                type: "runner",
-                filePath: ctx.filePath,
-                runnerId,
-                durationMs: 0,
-                status: "when_skipped",
-                diagnosticCount: 0,
-                semantic: "when_condition",
-            });
-            continue;
-        }
-        const projectRoot = ctx.projectRoot ?? ctx.cwd;
-        // Only post-write dispatches participate. Project scans and direct API
-        // callers must retain their existing synchronous semantics.
-        const observedTier = ctx.writeIndex === undefined
-            ? "inline"
-            : classifyObservedRunner(projectRoot, runner.id);
-        if (observedTier === "collect-later") {
-            const markedAtMs = Date.now();
-            const deferred = runRunner(ctx, runner, semantic).then((result) => {
-                const durationMs = Date.now() - markedAtMs;
-                const tier = observeRunnerLatency({
-                    projectRoot,
-                    runnerId: runner.id,
-                    durationMs,
-                    timedOut: result.failureKind === "timeout",
-                });
-                if (tier !== observedTier) {
-                    logLatency({
-                        type: "phase",
-                        filePath: ctx.filePath,
-                        phase: "runner_collect_later_tier_flip",
-                        durationMs: 0,
-                        metadata: {
-                            runnerId: runner.id,
-                            from: observedTier,
-                            to: tier,
-                            projectRoot,
-                        },
-                    });
-                }
-                if (tier === "collect-later") {
-                    incrementDegradationCount({
-                        kind: "runner-collect-later",
-                        subject: `${projectRoot}:${runner.id}`,
-                        reason: `observed ${durationMs}ms, threshold ${COLLECT_LATER_THRESHOLD_MS}ms`,
-                    });
-                }
-                logLatency({
-                    type: "runner",
-                    filePath: ctx.filePath,
-                    runnerId: runner.id,
-                    startedAt: new Date(markedAtMs).toISOString(),
-                    durationMs,
-                    status: result.status,
-                    diagnosticCount: result.diagnostics.length,
-                    semantic: result.semantic ?? semantic,
-                    metadata: { tier: "collect-later", delivered: "turn_end" },
-                });
-                return result;
-            });
-            deferRunnerFindings({
-                filePath: ctx.filePath,
-                cwd: ctx.cwd,
-                projectRoot,
-                runnerId: runner.id,
-                markedAtMs,
-                writeIndex: ctx.writeIndex,
-                promise: deferred,
-            });
-            // A deferred runner is still an observed runner. Keep it visible in
-            // both the edit latency report and the widget until its turn-end result
-            // replaces this pending state (#2122 F1).
-            latencies.push({
-                runnerId: runner.id,
-                startTime: runnerStart,
-                endTime: runnerStart,
-                durationMs: 0,
-                status: "pending",
-                diagnosticCount: 0,
-                semantic: semantic,
-            });
-            recordRunner(ctx.filePath, runner.id, "pending", 0, 0, ctx.writeIndex);
-            logLatency({
-                type: "runner",
-                filePath: ctx.filePath,
-                runnerId: runner.id,
-                durationMs: 0,
-                status: "pending",
-                diagnosticCount: 0,
-                semantic,
-                metadata: { tier: "collect-later", delivered: "turn_end" },
-            });
-            continue;
-        }
-        const result = await runRunner(ctx, runner, semantic);
-        onRunnerResult?.(runnerId, result);
-        const runnerEnd = Date.now();
-        const duration = runnerEnd - runnerStart;
-        const tier = ctx.writeIndex === undefined
-            ? "inline"
-            : observeRunnerLatency({
-                projectRoot,
-                runnerId: runner.id,
-                durationMs: duration,
-                timedOut: result.failureKind === "timeout",
-            });
-        if (tier !== observedTier) {
-            logLatency({
-                type: "phase",
-                filePath: ctx.filePath,
-                phase: "runner_collect_later_tier_flip",
-                durationMs: 0,
-                metadata: {
-                    runnerId: runner.id,
-                    from: observedTier,
-                    to: tier,
-                    projectRoot,
-                },
-            });
-        }
-        if (tier === "collect-later") {
-            incrementDegradationCount({
-                kind: "runner-collect-later",
-                subject: `${projectRoot}:${runner.id}`,
-                reason: `observed ${duration}ms, threshold ${COLLECT_LATER_THRESHOLD_MS}ms`,
-            });
-        }
-        // Runner definitions are a typed API, but embedders/plugins can still
-        // return untyped objects at runtime. Admit only the closed taxonomy and
-        // only on actual skips so free text cannot enter durable latency metadata.
-        const skipReason = result.status === "skipped" && isRunnerSkipReason(result.skipReason)
-            ? result.skipReason
-            : undefined;
-        latencies.push({
-            runnerId,
-            startTime: runnerStart,
-            endTime: runnerEnd,
-            durationMs: duration,
-            status: result.status,
-            diagnosticCount: result.diagnostics.length,
-            semantic: result.semantic ?? semantic,
-            ...(skipReason !== undefined && {
-                skipReason,
-            }),
-            ...(result.unconfirmedServerIds !== undefined && {
-                unconfirmedServerIds: result.unconfirmedServerIds,
-            }),
-            ...(result.deferredServerIds !== undefined && {
-                deferredServerIds: result.deferredServerIds,
-            }),
-        });
-        logLatency({
-            type: "runner",
-            filePath: ctx.filePath,
-            runnerId,
-            startedAt: new Date(runnerStart).toISOString(),
-            durationMs: duration,
-            status: result.status,
-            diagnosticCount: result.diagnostics.length,
-            semantic: result.semantic ?? semantic,
-            diagnostics: result.diagnostics.length > 0
-                ? result.diagnostics.map((d) => ({
-                    rule: d.rule,
-                    message: d.message.slice(0, 120),
-                    line: d.line,
-                    semantic: d.semantic,
-                }))
-                : undefined,
-            metadata: result.status === "failed" && result.failureKind
-                ? {
-                    failureKind: result.failureKind,
-                    failureMessage: result.failureMessage,
-                }
-                : skipReason
-                    ? { skipReason }
-                    : undefined,
-        });
-        recordRunner(ctx.filePath, runnerId, result.status, result.diagnostics.length, duration, ctx.writeIndex);
-        diagnostics.push(...result.diagnostics);
-        const resultSemantic = result.semantic ?? semantic;
-        if ((resultSemantic === "blocking" && result.diagnostics.length > 0) ||
-            result.diagnostics.some((d) => d.semantic === "blocking")) {
-            hadBlocker = true;
-        }
-        // mode:"fallback" — stop at first successful runner
-        if (group.mode === "fallback" && result.status === "succeeded") {
-            break;
-        }
-    }
-    return { diagnostics, latencies, hadBlocker };
-}
-// --- Main Dispatch Function ---
-export async function dispatchForFile(ctx, groups, registry, onRunnerResult) {
-    const _overallStart = Date.now();
-    if (ctx.fileRole === "generated") {
-        // The generated short-circuit (refs #2346): never ran before this fix
-        // for name-less machine-emitted files (a scraped/minified page has no
-        // `.min.js` pattern to catch it), so the classification now also has a
-        // content-shape tier. The skip is observable — one `dispatch_skipped_generated`
-        // phase record per file per process carrying the deciding evidence tier
-        // and the measured line-shape statistic. Deliberately NOT a degradation
-        // ledger entry: skipping a generated file is healthy behavior working
-        // as designed, and the ledger's bounded kind slots are reserved for
-        // genuine degradations (#2348 review F3 — the ledger precedent at the
-        // collect-later tier flip below records an actual capability loss).
-        const evidence = ctx.generatedEvidence;
-        const lineShapeMean = ctx.generatedLineShapeMean;
-        if (!generatedSkipRecorded.has(ctx.filePath)) {
-            generatedSkipRecorded.add(ctx.filePath);
-            logLatency({
-                type: "phase",
-                filePath: ctx.filePath,
-                phase: "dispatch_skipped_generated",
-                durationMs: 0,
-                metadata: {
-                    evidence: evidence ?? "unknown",
-                    ...(lineShapeMean !== undefined && { lineShapeMean }),
-                },
-            });
-        }
-        return {
-            diagnostics: [],
-            blockers: [],
-            warnings: [],
-            baselineWarningCount: 0,
-            fixed: [],
-            resolvedCount: 0,
-            output: "",
-            blockerOutput: "",
-            hasBlockers: false,
-        };
-    }
-    const allDiagnostics = [];
-    let stopped = false;
-    const runnerLatencies = [];
-    // Debug logging goes to latency log only (not console - avoid noise)
-    const allRunnerIds = groups.flatMap((g) => g.runnerIds);
-    logLatency({
-        type: "phase",
-        filePath: ctx.filePath,
-        phase: "dispatch_start",
-        durationMs: 0,
-        metadata: {
-            groupCount: groups.length,
-            kind: ctx.kind,
-            runners: allRunnerIds.join(","),
-        },
-    });
-    // Run all groups in parallel — they are independent and don't depend on
-    // each other's results. Within each group, mode:"fallback" semantics are
-    // preserved (sequential first-success). Results are merged in original
-    // group order so output is deterministic.
-    const groupResults = await Promise.all(groups.map((group) => runGroup(ctx, group, registry, onRunnerResult)));
-    // Count baseline warnings before filtering (for delta count display)
-    //
-    // #2489: `ctx.filePath` is `createDispatchContext`'s `normalizedFilePath`
-    // (`resolveRunnerPath` -> `path.resolve` -> `normalizeMapKey`), which is
-    // ALWAYS absolute — the #2016 invariant this seam already relies on. A
-    // prior relative-path fallback key (`session.baseline.<path relative to
-    // ctx.cwd>`) existed here for self-consistency across the abs/rel read
-    // and write, but on `ctx.facts` singletons that outlive one project root
-    // (`dispatch/integration.ts`'s module-scope `sessionFacts`, reached by
-    // the warm `pilens_analyze` MCP route across any number of cwds) two
-    // different projects dispatching files that share a relative path (e.g.
-    // both have `src/index.ts`) collided on that fallback key, so project
-    // B's first-ever delta baseline read could silently return project A's
-    // stored diagnostics. Since the absolute key can never collide across
-    // projects (an absolute path is unique on the filesystem) and both
-    // writes always land together, the relative fallback is both unneeded
-    // and the sole source of the collision — removed rather than cwd-scoped
-    // (`cwd` is not a stable identity across this seam's callers, #2494).
-    //
-    // #2489 round 2: the #2016 invariant above is enforced by
-    // `createDispatchContext`, the sole real constructor, but nothing enforced
-    // it AT THIS SEAM — a hand-built `DispatchContext` (test scaffolding, a
-    // future caller) carrying a relative `filePath` would silently key the
-    // baseline under a value the constructor could never produce, and every
-    // subsequent same-session dispatch of the real (absolute-keyed) file would
-    // read back `undefined`, never the collision itself. Mirrors the sibling
-    // guard `recordEntitySnapshotDiff` already enforces for the entity-snapshot
-    // seam (`clients/review-graph/service.ts`, #2477): reject visibly and skip
-    // the baseline read/write for this dispatch rather than compute one under
-    // an unreachable key.
-    const baselinePathSafe = isAbsoluteFilePath(ctx.filePath);
-    if (ctx.deltaMode && !baselinePathSafe) {
-        recordDegradationOnce({
-            kind: "dispatch-non-absolute-baseline-path",
-            subject: ctx.filePath,
-            reason: "dispatchForFile requires an absolute ctx.filePath to key the delta baseline (refs #2489)",
-        });
-    }
-    const baselineAbsKey = `session.baseline.${ctx.filePath}`;
-    const previousBaseline = ctx.deltaMode && baselinePathSafe
-        ? ctx.facts.getBoundedSessionFact(baselineAbsKey)
-        : undefined;
-    const baselineWarnings = previousBaseline?.filter((d) => d.semantic === "warning" || d.semantic === "none");
-    const baselineWarningCount = baselineWarnings?.length ?? 0;
-    for (const { diagnostics: groupDiags, latencies, hadBlocker, } of groupResults) {
-        runnerLatencies.push(...latencies);
-        allDiagnostics.push(...groupDiags);
-        if (hadBlocker)
-            stopped = true;
-    }
-    // Apply delta mode ONCE across the full diagnostic set.
-    // This avoids partial-baseline corruption when processing multiple groups.
-    const dedupedDiagnostics = dedupeOverlappingDiagnostics(allDiagnostics);
-    const fileContent = ctx.facts.getFileFact(ctx.filePath, "file.content") ?? "";
-    // Project rule policy (`.pi-lens.json` `rules.<id>.disable`/`select`).
-    //
-    // Resolved from `ctx.projectRoot` (falling back to `ctx.cwd`), NOT
-    // `ctx.projectConfig` — `ctx.projectConfig` is loaded from the nested
-    // LANGUAGE root (`resolveLanguageRootForFile`), so in a monorepo where a
-    // package directory has its own `.pi-lens.json`, `discoverPiLensProjectConfig`'s
-    // upward walk stops there and never sees a repo-root policy.
-    // `lens_diagnostics` loads its policy map from `runtime.projectRoot`; using
-    // the same root here keeps the two surfaces in agreement. `ctx.projectConfig`
-    // itself is untouched — thresholds and mutation flags keep their existing
-    // language-root resolution.
-    const rulePolicy = rulePolicyMapFromConfig(loadPiLensProjectConfig(ctx.projectRoot ?? ctx.cwd).rules);
-    // The output-only filter pipeline: LSP/docker overlap suppression + inline
-    // `pi-lens-ignore` + agent/user dispositions + project rule policy. Applied
-    // AFTER dedupe so the pi renderer, widget, and delta all see one filtered
-    // set. #690: dispositions drop false-positive/suppress marks and anything
-    // deferred this session (flagged marks stay; lens_diagnostics tags them at
-    // render time). #1030: the disposition anchor + store MUST key off the
-    // PROJECT ROOT, not ctx.cwd — the mark tool writes dispositions under
-    // runtime.projectRoot, so reading from ctx.cwd (the nested language root in
-    // a monorepo) opened a different diagnostic-dispositions.json and silently
-    // no-op'd every mark under a nested marker.
-    //
-    // #1087: "silencing is not fixing" applies to the WHOLE class, not just the
-    // policy member. This single pipeline is applied identically to the live set
-    // AND (below) to the delta baseline, so a finding persistently dropped by
-    // ANY layer — overlap, inline suppression, disposition, or policy — is
-    // absent from both sides of the delta and never oscillates into `fixed`
-    // (which `trackAgentFixed` would otherwise inflate forever). The stored
-    // baseline remains the unfiltered `dedupedDiagnostics`, so editing a
-    // project's policy/suppressions never resets or corrupts the user-authored
-    // delta baseline — the filtering is re-derived on read for both sides.
-    const applyOutputFilters = (diags) => {
-        const dockerOverlap = suppressTrivyConfigDockerOverlap(diags);
-        const overlap = suppressLintOverlapsWithLsp(dockerOverlap);
-        const inline = applyInlineSuppressions(overlap, fileContent);
-        const disposition = applyDispositions(inline, ctx.projectRoot ?? ctx.cwd, ctx.filePath, fileContent);
-        return applyRulePolicy(disposition, rulePolicy);
-    };
-    let visibleDiagnostics = applyOutputFilters(dedupedDiagnostics);
-    let resolvedCount = 0;
-    if (ctx.deltaMode && previousBaseline) {
-        // Silencing a rule is not fixing it. The stored baseline is deliberately
-        // unfiltered (below), so compare against a fully-filtered view of it
-        // through the SAME pipeline — otherwise every persistently-suppressed
-        // finding sits in `fixed` on every dispatch and inflates the agent's
-        // resolved tally (trackAgentFixed) forever. Only `fixed` changes:
-        // filtered-out ids are absent from `after` either way.
-        const filtered = filterDelta(visibleDiagnostics, applyOutputFilters(previousBaseline), (d) => d.id);
-        visibleDiagnostics = promoteDeltaUnusedToBlockers(filtered.new);
-        resolvedCount = filtered.fixed.length;
-    }
-    // Persist full current snapshot for next run (not delta-filtered subset).
-    // Gated on `baselinePathSafe` too: a non-absolute `ctx.filePath` already
-    // skipped the read above (the degradation record fired there), and writing
-    // under that same unreachable key here would just seed a baseline no real
-    // reader (always absolute) could ever read back.
-    if (ctx.deltaMode && baselinePathSafe) {
-        ctx.facts.setBoundedSessionFact(baselineAbsKey, [...dedupedDiagnostics]);
-    }
-    // Categorize results
-    const blockers = visibleDiagnostics.filter((d) => d.semantic === "blocking");
-    const warnings = visibleDiagnostics.filter((d) => d.semantic === "warning" || d.semantic === "none");
-    const fixedItems = visibleDiagnostics.filter((d) => d.semantic === "fixed");
-    // Append fixed and fixable diagnostics to the persistent worklog, attributed
-    // to the runtime's active model/provider when known (#1448).
-    const worklogIdentity = {
-        model: ctx.telemetryModel,
-        provider: ctx.telemetryProvider,
-    };
-    if (fixedItems.length > 0) {
-        import("../fix-worklog.js")
-            .then(({ appendToWorklog }) => {
-            appendToWorklog(ctx.cwd, fixedItems, true, worklogIdentity);
-        })
-            .catch(() => { });
-    }
-    const fixableWarnings = warnings.filter((d) => d.fixable);
-    if (fixableWarnings.length > 0) {
-        import("../fix-worklog.js")
-            .then(({ appendToWorklog }) => {
-            appendToWorklog(ctx.cwd, fixableWarnings, false, worklogIdentity);
-        })
-            .catch(() => { });
-    }
-    const inlineBlockers = blockers;
-    const inlineFixed = fixedItems;
-    const coverageNotice = buildCoverageNotice(ctx, runnerLatencies);
-    // Format output — only blocking issues shown inline
-    // Warnings tracked but not shown (noise) — surfaced via lens_diagnostics
-    const blockerOutput = formatDiagnostics(inlineBlockers, "blocking");
-    let output = blockerOutput;
-    output += formatDiagnostics(inlineFixed, "fixed");
-    if (coverageNotice) {
-        output += formatDiagnostics([coverageNotice], "warning", 1);
-        warnings.push(coverageNotice);
-    }
-    const pendingRunners = runnerLatencies
-        .filter((runner) => runner.status === "pending")
-        .map((runner) => runner.runnerId);
-    if (pendingRunners.length > 0) {
-        output += `\n⏳ Pending runners (reported at turn end): ${pendingRunners.join(", ")}\n`;
-    }
-    // Generate and store latency report
-    const overallEnd = Date.now();
-    const latencyReport = {
-        filePath: ctx.filePath,
-        fileKind: ctx.kind,
-        overallStartMs: _overallStart,
-        overallEndMs: overallEnd,
-        totalDurationMs: overallEnd - _overallStart,
-        runners: runnerLatencies,
-        stoppedEarly: stopped,
-        totalDiagnostics: visibleDiagnostics.length,
-        blockers: blockers.length,
-        warnings: warnings.length,
-    };
-    // Store for later analysis
-    latencyReports.push(latencyReport);
-    // Keep only last 100 reports to prevent memory bloat
-    if (latencyReports.length > 100) {
-        latencyReports.shift();
-    }
-    // Runner latencies already logged immediately after execution (line ~329)
-    // The runnerLatencies array is stored in latencyReport for aggregate analysis
-    // No need to log again here - would create duplicates in the log
-    // Log summary to latency log only (not console - avoid noise)
-    const sumMs = runnerLatencies.reduce((s, r) => s + r.durationMs, 0);
-    const wallClockMs = latencyReport.totalDurationMs;
-    logLatency({
-        type: "tool_result",
-        filePath: ctx.filePath,
-        durationMs: wallClockMs,
-        wallClockMs,
-        sumMs,
-        parallelGainMs: Math.max(0, sumMs - wallClockMs),
-        result: "dispatch_complete",
-        metadata: {
-            runners: runnerLatencies.map((r) => ({
-                id: r.runnerId,
-                startedAt: new Date(r.startTime).toISOString(),
-                duration: r.durationMs,
-                status: r.status,
-            })),
-            totalDiagnostics: visibleDiagnostics.length,
-            blockers: blockers.length,
-            // #2489 round 2: distinguishes a delta baseline HIT (a prior snapshot
-            // was read for this file this session) from a MISS (first dispatch, or
-            // the non-absolute-path guard above skipped the read) — `dispatch_start`
-            // (above, `metadata: { groupCount, kind, runners }`) fires before the
-            // baseline read ever runs and cannot carry either value.
-            baselineHit: previousBaseline !== undefined,
-            baselineWarningCount,
-        },
-    });
-    return {
-        diagnostics: visibleDiagnostics,
-        blockers,
-        warnings,
-        baselineWarningCount,
-        fixed: fixedItems,
-        resolvedCount,
-        output,
-        blockerOutput,
-        hasBlockers: blockers.length > 0,
-    };
-}
-// --- Run Single Runner ---
-/** Maximum wall-clock time a single runner may take before we abort it. */
-const RUNNER_TIMEOUT_MS = RUNTIME_CONFIG.dispatch.runnerTimeoutMs;
-function looksLikeDiagnosticCodePath(value) {
-    if (!value)
-        return false;
-    const text = value.trim();
-    if (!text)
-        return false;
-    const base = path.basename(text.replace(/\\/g, "/"));
-    if (/^lsp:\d+(?::\d+)?$/i.test(text) || /^lsp:\d+(?::\d+)?$/i.test(base)) {
-        return true;
-    }
-    if (/^similarity[-:]/i.test(text) || /^similarity[-:]/i.test(base)) {
-        return true;
-    }
-    if (/^[a-z-]+:\d+(?::\d+)?$/i.test(text) ||
-        /^[a-z-]+:\d+(?::\d+)?$/i.test(base)) {
-        return true;
-    }
-    return false;
-}
-function normalizeDiagnosticFilePath(ctx, rawPath) {
-    if (typeof rawPath === "string" && looksLikeDiagnosticCodePath(rawPath)) {
-        ctx.log(`runner path normalization: ignored diagnostic code-like path '${rawPath}', using current file`);
-        return resolveRunnerPath(ctx.cwd, ctx.filePath);
-    }
-    return resolveRunnerPath(ctx.cwd, rawPath || ctx.filePath);
-}
-async function runRunner(ctx, runner, defaultSemantic) {
-    const timeoutMs = Math.max(runner.timeoutMs ?? RUNNER_TIMEOUT_MS, getRunnerTimeoutFloorMs());
-    let timer;
-    // #1723: mark this runner as the in-flight phase for the WHOLE race below,
-    // not just the runner's own promise — a synchronous CPU hog inside
-    // `runner.run` (ast-grep-napi, tree-sitter, …) blocks the event loop before
-    // it ever gets to log its own completion, so `recentPhases` (which only
-    // records FINISHED phases) can't name it. `phaseFinished` clears on every
-    // exit from this race (success, runner error, or timeout) via try/finally,
-    // so a `loop_block` sampled after this call returns never sees a stale
-    // pointer. See `phaseStarted`'s doc comment for the identity-token
-    // reasoning and the cost note (negligible next to a runner invocation).
-    const phaseToken = phaseStarted(runner.id);
-    try {
-        const result = await Promise.race([
-            runner.run(ctx).finally(() => clearTimeout(timer)),
-            new Promise((_, reject) => {
-                timer = setTimeout(() => reject(new Error(`Runner ${runner.id} timed out after ${timeoutMs}ms`)), timeoutMs);
-            }),
-        ]);
-        const diagnostics = result.diagnostics.map((d) => ({
-            ...d,
-            filePath: normalizeDiagnosticFilePath(ctx, d.filePath),
-        }));
-        return {
-            ...result,
-            diagnostics,
-            semantic: result.semantic ?? defaultSemantic,
-        };
-    }
-    catch (error) {
-        clearTimeout(timer);
-        ctx.log(`Runner ${runner.id} failed: ${error}`);
-        const message = error instanceof Error ? error.message : String(error);
-        return {
-            status: "failed",
-            diagnostics: [],
-            semantic: defaultSemantic,
-            failureKind: message.includes("timed out") ? "timeout" : "exception",
-            failureMessage: message.slice(0, 200),
-        };
-    }
-    finally {
-        phaseFinished(phaseToken);
-    }
-}
-// --- Simple Integration Helper ---