@event4u/agent-config 9.9.0 → 9.10.0

package/src/scripts/preamble_byte_census.ts

@@ -0,0 +1,574 @@
+#!/usr/bin/env node
+/**
+ * Preamble byte census — attributes the measured subagent cold-start payload
+ * to named sources (`road-to-cache-economy.md` Phase 3,
+ * steps 1-2): user-scope rules, project-scope rules, the CLAUDE.md
+ * hierarchy (project + user + `@`-imports + `CLAUDE.local.md`), the global
+ * user `profile.md` (road-to-global-user-memory), and the preloaded skills
+ * catalog (name + `description` frontmatter, the shape the host actually
+ * injects — verified against this repo's own skill-catalog listing).
+ *
+ * Everything NOT independently file-measurable — the host's raw
+ * tool-definition JSON and the per-spawn dispatch-prompt text — is reported
+ * as ONE labelled residual: measured_cold_start_median minus the sum of the
+ * file-measurable buckets above. That is disclosed, not hidden: no local
+ * transcript or file carries the request's system/tool payload (verified —
+ * `~/.claude/projects/**\/*.jsonl` records only `message.usage` counts and
+ * response content; a raw scan of a real transcript file's top-level keys
+ * turned up no `system`, `tools`, or prompt-body field). Because the
+ * residual is defined as the gap, the "buckets sum to the measured median"
+ * check is trivially satisfied BY CONSTRUCTION once a residual exists — this
+ * is stated plainly in the report rather than presented as an independent
+ * confirmation. The independently interesting number is what SHARE of the
+ * median the file-measurable buckets cover on their own.
+ *
+ * Also reports the MODELLED (never "measured") reduction from removing the
+ * duplicate-scope copy the C-2 census already finds — see
+ * `_lib/duplicate_scope_census.ts`, reused here rather than reimplemented.
+ *
+ * Class A (per the ADR-124 no-runtime-boundary contract): in-process,
+ * per-invocation, no socket, no daemon, no network. Reads local files
+ * (rules, CLAUDE.md hierarchy, skill frontmatter, transcripts) only.
+ *
+ * Usage:
+ *   ./scripts-run src/scripts/preamble_byte_census [--format text|json]
+ *     [--repo-root <path>] [--user-home <path>] [--user-rules-dir <path>]
+ *     [--project-rules-dir <path>] [--skills-dir <path>]
+ *     [--root <path>] [--max-age-days <n>]
+ */
+import * as fs from 'node:fs';
+import { homedir } from 'node:os';
+import * as path from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+
+import * as YAML from 'yaml';
+
+import { resolveGlobalProfilePath } from './_lib/agent_user_profile.js';
+import { DEFAULT_PROJECTS_ROOT, scanTranscripts } from './_lib/cc_transcript.js';
+import { censusDuplicateScope } from './_lib/duplicate_scope_census.js';
+import { computeColdStarts } from './cache_realization_report.js';
+
+const _HERE = fileURLToPath(import.meta.url);
+const REPO_ROOT = path.resolve(path.dirname(_HERE), '..', '..');
+const CHARS_PER_TOKEN = 4;
+
+// A sum within this tolerance of the measured median counts as "reconciled" —
+// mirrors the roadmap's own ±10% verify line. See the module doc comment for
+// why this is trivially satisfied once the residual bucket exists.
+const SUM_TOLERANCE_PCT = 0.1;
+
+function tokensFromChars(chars: number): number {
+    return chars / CHARS_PER_TOKEN;
+}
+
+function readCharsIfExists(p: string): number {
+    try {
+        return fs.statSync(p).size;
+    } catch {
+        return 0;
+    }
+}
+
+// ── source 1/2: rule directories ────────────────────────────────────────
+
+export interface RuleDirCensus {
+    files: number;
+    chars: number;
+}
+
+export function censusRuleDir(dir: string): RuleDirCensus {
+    if (!fs.existsSync(dir)) return { files: 0, chars: 0 };
+    let files = 0;
+    let chars = 0;
+    for (const name of fs.readdirSync(dir)) {
+        if (!name.endsWith('.md')) continue;
+        try {
+            chars += fs.statSync(path.join(dir, name)).size;
+            files += 1;
+        } catch {
+            // Unreadable entry — skip rather than fail the whole census.
+        }
+    }
+    return { files, chars };
+}
+
+export interface PerRuleCost {
+    file: string;
+    chars: number;
+    tokens_estimate: number;
+}
+
+/** Top `limit` `.md` files in `dir` by byte size, descending — the per-rule per-spawn cost list (step 2). */
+// A candidate list, not a verdict: routing a rule to dormancy goes through
+// _lib/compile_time_toggles.ts (see its "dormancy routing" header) and needs an
+// output-side bench first. Cost alone never justifies a flip.
+export function topRulesByCost(dir: string, limit = 10): PerRuleCost[] {
+    if (!fs.existsSync(dir)) return [];
+    const entries: PerRuleCost[] = [];
+    for (const name of fs.readdirSync(dir)) {
+        if (!name.endsWith('.md')) continue;
+        let chars: number;
+        try {
+            chars = fs.statSync(path.join(dir, name)).size;
+        } catch {
+            continue;
+        }
+        entries.push({ file: name, chars, tokens_estimate: tokensFromChars(chars) });
+    }
+    entries.sort((a, b) => b.chars - a.chars || a.file.localeCompare(b.file));
+    return entries.slice(0, limit);
+}
+
+// ── source 3: CLAUDE.md hierarchy (project + user + @-imports + local) ──
+
+export interface ClaudeMdCensus {
+    project_claude_md_chars: number;
+    project_claude_md_present: boolean;
+    project_claude_local_md_chars: number;
+    project_claude_local_md_present: boolean;
+    user_claude_md_chars: number;
+    user_claude_md_present: boolean;
+    /** Files pulled in by a top-level `@<path>` import line in the user CLAUDE.md. */
+    user_imports: { file: string; chars: number }[];
+}
+
+/** `@<path>` import lines, one per line, at the START of a line (Claude Code's own import syntax). */
+const IMPORT_LINE_RE = /^@(\S+)\s*$/;
+
+function parseClaudeMdImports(claudeMdPath: string): { file: string; chars: number }[] {
+    let text: string;
+    try {
+        text = fs.readFileSync(claudeMdPath, 'utf-8');
+    } catch {
+        return [];
+    }
+    const base = path.dirname(claudeMdPath);
+    const out: { file: string; chars: number }[] = [];
+    for (const line of text.split('\n')) {
+        const m = IMPORT_LINE_RE.exec(line.trim());
+        if (!m) continue;
+        const rel = m[1] as string;
+        const resolved = path.isAbsolute(rel) ? rel : path.join(base, rel);
+        const chars = readCharsIfExists(resolved);
+        if (chars > 0) out.push({ file: rel, chars });
+    }
+    return out;
+}
+
+export function censusClaudeMdHierarchy(repoRoot: string, userHome: string): ClaudeMdCensus {
+    const projectClaudeMd = path.join(repoRoot, 'CLAUDE.md');
+    const projectClaudeLocalMd = path.join(repoRoot, 'CLAUDE.local.md');
+    const userClaudeMd = path.join(userHome, '.claude', 'CLAUDE.md');
+
+    return {
+        project_claude_md_chars: readCharsIfExists(projectClaudeMd),
+        project_claude_md_present: fs.existsSync(projectClaudeMd),
+        project_claude_local_md_chars: readCharsIfExists(projectClaudeLocalMd),
+        project_claude_local_md_present: fs.existsSync(projectClaudeLocalMd),
+        user_claude_md_chars: readCharsIfExists(userClaudeMd),
+        user_claude_md_present: fs.existsSync(userClaudeMd),
+        user_imports: parseClaudeMdImports(userClaudeMd),
+    };
+}
+
+// ── source 4: global user profile.md (road-to-global-user-memory) ───────
+
+export interface GlobalProfileCensus {
+    present: boolean;
+    chars: number;
+    path: string | null;
+}
+
+/**
+ * `resolveGlobalProfilePath` returns `null` when the layer hasn't landed /
+ * isn't written yet — treated as 0, per the roadmap's own honesty caveat.
+ * `env` is injectable so tests never touch this machine's real
+ * `~/.event4u/agent-config` — see `agent_user_profile.ts`.
+ */
+export function censusGlobalProfile(env?: Parameters<typeof resolveGlobalProfilePath>[0]): GlobalProfileCensus {
+    const p = resolveGlobalProfilePath(env);
+    if (p === null) return { present: false, chars: 0, path: null };
+    return { present: true, chars: readCharsIfExists(p), path: p };
+}
+
+// ── source 5: preloaded skills catalog (name + description) ─────────────
+
+export interface SkillsCatalogCensus {
+    skills: number;
+    chars: number;
+}
+
+interface RawSkillFrontmatter {
+    name?: unknown;
+    description?: unknown;
+}
+
+function parseFrontmatter(text: string): RawSkillFrontmatter {
+    if (!text.startsWith('---')) return {};
+    const end = text.indexOf('\n---', 3);
+    if (end === -1) return {};
+    const block = text.slice(3, end);
+    try {
+        const parsed: unknown = YAML.parse(block);
+        return typeof parsed === 'object' && parsed !== null ? (parsed as RawSkillFrontmatter) : {};
+    } catch {
+        return {};
+    }
+}
+
+/**
+ * Sums `- <name>: <description>\n` across every `SKILL.md` under `skillsDir`
+ * — the shape the host's own skill-catalog listing uses (verified against
+ * this session's own system-reminder catalog). This is the description-text
+ * component only; the exact on-wire bullet/wrapping format the host applies
+ * on top is not itself file-measurable, so this is a lower-bound estimate,
+ * stated as such wherever it is reported.
+ */
+export function censusSkillsCatalog(skillsDir: string): SkillsCatalogCensus {
+    if (!fs.existsSync(skillsDir)) return { skills: 0, chars: 0 };
+    let skills = 0;
+    let chars = 0;
+    for (const entry of fs.readdirSync(skillsDir, { withFileTypes: true })) {
+        if (!entry.isDirectory()) continue;
+        const skillMd = path.join(skillsDir, entry.name, 'SKILL.md');
+        let text: string;
+        try {
+            text = fs.readFileSync(skillMd, 'utf-8');
+        } catch {
+            continue;
+        }
+        const fm = parseFrontmatter(text);
+        const name = typeof fm.name === 'string' && fm.name.length > 0 ? fm.name : entry.name;
+        const description = typeof fm.description === 'string' ? fm.description : '';
+        chars += `- ${name}: ${description}\n`.length;
+        skills += 1;
+    }
+    return { skills, chars };
+}
+
+// ── report assembly ───────────────────────────────────────────────────────
+
+export interface ByteCensusSource {
+    name: string;
+    files: number;
+    chars: number;
+    tokens_estimate: number;
+    /** Estimate vs residual — see module doc comment. */
+    provenance: 'measured_local_file' | 'residual';
+    /** How this figure was derived — mandatory for anything not a plain file-size sum. */
+    basis: string;
+}
+
+export interface ByteCensus {
+    sources: ByteCensusSource[];
+    measurable_tokens_total: number;
+    /** Includes the residual bucket — always ≈ measured_cold_start_median by construction (see module doc comment). */
+    grand_total_tokens: number;
+    measured_cold_start_median: number | null;
+    cold_start_legs: number;
+    /** measurable_tokens_total / measured_cold_start_median — the honest, non-tautological number. */
+    measurable_share_of_median: number | null;
+    within_tolerance: boolean | null;
+    top_rules_by_cost: PerRuleCost[];
+    duplicate_scope: {
+        evaluable: boolean;
+        reason?: string | undefined;
+        shared_filenames: number;
+        duplicate_tokens_per_spawn: number;
+    };
+    /** MODELLED, never measured — see module doc comment and the `modelled` label on every field here. */
+    modelled_duplicate_removal: {
+        applicable: boolean;
+        reason?: string | undefined;
+        modelled_new_median: number | null;
+        modelled_reduction_pct: number | null;
+    };
+}
+
+export interface Options {
+    format: 'text' | 'json';
+    repoRoot: string;
+    userHome: string;
+    userRulesDir: string;
+    projectRulesDir: string;
+    skillsDir: string;
+    root: string;
+    maxAgeDays: number;
+}
+
+function defaultOptions(): Options {
+    const userHome = homedir();
+    return {
+        format: 'text',
+        repoRoot: REPO_ROOT,
+        userHome,
+        userRulesDir: path.join(userHome, '.claude', 'rules'),
+        projectRulesDir: path.join(REPO_ROOT, 'dist', 'agent-src', 'rules'),
+        skillsDir: path.join(REPO_ROOT, 'dist', 'agent-src', 'skills'),
+        root: DEFAULT_PROJECTS_ROOT,
+        maxAgeDays: 14,
+    };
+}
+
+export function parseArgs(argv: string[]): Options {
+    const opts = defaultOptions();
+    for (let i = 0; i < argv.length; i++) {
+        const a = argv[i];
+        if (a === undefined) continue;
+        if (a === '--format') opts.format = (argv[++i] as Options['format']) ?? opts.format;
+        else if (a.startsWith('--format=')) opts.format = a.slice('--format='.length) as Options['format'];
+        else if (a === '--repo-root') opts.repoRoot = argv[++i] ?? opts.repoRoot;
+        else if (a.startsWith('--repo-root=')) opts.repoRoot = a.slice('--repo-root='.length);
+        else if (a === '--user-home') opts.userHome = argv[++i] ?? opts.userHome;
+        else if (a.startsWith('--user-home=')) opts.userHome = a.slice('--user-home='.length);
+        else if (a === '--user-rules-dir') opts.userRulesDir = argv[++i] ?? opts.userRulesDir;
+        else if (a.startsWith('--user-rules-dir=')) opts.userRulesDir = a.slice('--user-rules-dir='.length);
+        else if (a === '--project-rules-dir') opts.projectRulesDir = argv[++i] ?? opts.projectRulesDir;
+        else if (a.startsWith('--project-rules-dir=')) opts.projectRulesDir = a.slice('--project-rules-dir='.length);
+        else if (a === '--skills-dir') opts.skillsDir = argv[++i] ?? opts.skillsDir;
+        else if (a.startsWith('--skills-dir=')) opts.skillsDir = a.slice('--skills-dir='.length);
+        else if (a === '--root') opts.root = argv[++i] ?? opts.root;
+        else if (a.startsWith('--root=')) opts.root = a.slice('--root='.length);
+        else if (a === '--max-age-days') opts.maxAgeDays = Number(argv[++i]) || opts.maxAgeDays;
+        else if (a.startsWith('--max-age-days=')) opts.maxAgeDays = Number(a.slice('--max-age-days='.length)) || opts.maxAgeDays;
+    }
+    if (opts.format !== 'text' && opts.format !== 'json') opts.format = 'text';
+    return opts;
+}
+
+/**
+ * Assembles the full census from already-computed measurements — pure, so
+ * tests can supply synthetic figures without touching real transcripts or
+ * this machine's `~/.claude`.
+ */
+export function buildByteCensus(opts: {
+    userRules: RuleDirCensus;
+    projectRules: RuleDirCensus;
+    claudeMd: ClaudeMdCensus;
+    globalProfile: GlobalProfileCensus;
+    skillsCatalog: SkillsCatalogCensus;
+    topRules: PerRuleCost[];
+    duplicateScope: ReturnType<typeof censusDuplicateScope>;
+    measuredColdStartMedian: number | null;
+    coldStartLegs: number;
+}): ByteCensus {
+    const claudeMdChars =
+        opts.claudeMd.project_claude_md_chars +
+        opts.claudeMd.project_claude_local_md_chars +
+        opts.claudeMd.user_claude_md_chars +
+        opts.claudeMd.user_imports.reduce((s, i) => s + i.chars, 0);
+    const claudeMdFiles =
+        (opts.claudeMd.project_claude_md_present ? 1 : 0) +
+        (opts.claudeMd.project_claude_local_md_present ? 1 : 0) +
+        (opts.claudeMd.user_claude_md_present ? 1 : 0) +
+        opts.claudeMd.user_imports.length;
+
+    const sources: ByteCensusSource[] = [
+        {
+            name: 'user-scope rules (~/.claude/rules/*.md)',
+            files: opts.userRules.files,
+            chars: opts.userRules.chars,
+            tokens_estimate: tokensFromChars(opts.userRules.chars),
+            provenance: 'measured_local_file',
+            basis: 'Σ byte size of every .md file in the user-scope rules directory, chars/4.',
+        },
+        {
+            name: 'project-scope rules (dist/agent-src/rules/*.md)',
+            files: opts.projectRules.files,
+            chars: opts.projectRules.chars,
+            tokens_estimate: tokensFromChars(opts.projectRules.chars),
+            provenance: 'measured_local_file',
+            basis: 'Σ byte size of every .md file in the project-scope rules directory, chars/4.',
+        },
+        {
+            name: 'CLAUDE.md hierarchy (project + project-local + user + user @-imports)',
+            files: claudeMdFiles,
+            chars: claudeMdChars,
+            tokens_estimate: tokensFromChars(claudeMdChars),
+            provenance: 'measured_local_file',
+            basis:
+                'Byte size of project CLAUDE.md + project CLAUDE.local.md (if present) + user ~/.claude/CLAUDE.md ' +
+                '+ every file pulled in by a top-level "@<path>" import line in the user CLAUDE.md, chars/4.',
+        },
+        {
+            name: 'global user profile.md (road-to-global-user-memory)',
+            files: opts.globalProfile.present ? 1 : 0,
+            chars: opts.globalProfile.chars,
+            tokens_estimate: tokensFromChars(opts.globalProfile.chars),
+            provenance: 'measured_local_file',
+            basis: opts.globalProfile.present
+                ? `Byte size of ${opts.globalProfile.path as string}, chars/4.`
+                : 'Layer not written yet on this machine (resolveGlobalProfilePath() returned null) — treated as 0, per the roadmap\'s own honesty caveat.',
+        },
+        {
+            name: 'preloaded skills catalog (name + description, dist/agent-src/skills)',
+            files: opts.skillsCatalog.skills,
+            chars: opts.skillsCatalog.chars,
+            tokens_estimate: tokensFromChars(opts.skillsCatalog.chars),
+            provenance: 'measured_local_file',
+            basis:
+                'Σ "- <name>: <description>\\n" over every SKILL.md frontmatter, chars/4 — the description-text ' +
+                'component only; the host\'s exact on-wire catalog format is not independently file-measurable.',
+        },
+    ];
+
+    const measurableTokensTotal = sources.reduce((s, src) => s + src.tokens_estimate, 0);
+    const median = opts.measuredColdStartMedian;
+    const residualTokens = median !== null ? Math.max(0, median - measurableTokensTotal) : null;
+
+    if (residualTokens !== null) {
+        sources.push({
+            name: 'tool definitions + dispatch prompt (residual)',
+            files: 0,
+            chars: 0,
+            tokens_estimate: residualTokens,
+            provenance: 'residual',
+            basis:
+                'measured_cold_start_median minus the sum of the file-measurable buckets above. No local transcript ' +
+                'or file carries the request\'s raw tool-definition JSON or the per-spawn dispatch-prompt text ' +
+                '(verified: ~/.claude/projects/**/*.jsonl records only message.usage counts and response content, ' +
+                'never the request system/tool payload) — this bucket exists so the gap is named rather than silently dropped.',
+        });
+    }
+
+    const grandTotal = sources.reduce((s, src) => s + src.tokens_estimate, 0);
+    const measurableShare = median !== null && median > 0 ? measurableTokensTotal / median : null;
+    const withinTolerance = median !== null ? Math.abs(grandTotal - median) / Math.max(1, median) <= SUM_TOLERANCE_PCT : null;
+
+    const dup = opts.duplicateScope;
+    let modelledNewMedian: number | null = null;
+    let modelledReductionPct: number | null = null;
+    let modelledApplicable = false;
+    let modelledReason: string | undefined;
+    if (median === null) {
+        modelledReason = 'no measured cold-start median available in this window';
+    } else if (!dup.evaluable) {
+        modelledReason = dup.reason ?? 'duplicate-scope census not evaluable';
+    } else {
+        modelledApplicable = true;
+        modelledNewMedian = Math.max(0, median - dup.duplicate_chars / CHARS_PER_TOKEN);
+        modelledReductionPct = median > 0 ? (median - modelledNewMedian) / median : 0;
+    }
+
+    return {
+        sources,
+        measurable_tokens_total: measurableTokensTotal,
+        grand_total_tokens: grandTotal,
+        measured_cold_start_median: median,
+        cold_start_legs: opts.coldStartLegs,
+        measurable_share_of_median: measurableShare,
+        within_tolerance: withinTolerance,
+        top_rules_by_cost: opts.topRules,
+        duplicate_scope: {
+            evaluable: dup.evaluable,
+            reason: dup.reason,
+            shared_filenames: dup.shared_filenames.length,
+            duplicate_tokens_per_spawn: dup.evaluable ? dup.duplicate_chars / CHARS_PER_TOKEN : 0,
+        },
+        modelled_duplicate_removal: {
+            applicable: modelledApplicable,
+            reason: modelledReason,
+            modelled_new_median: modelledNewMedian,
+            modelled_reduction_pct: modelledReductionPct,
+        },
+    };
+}
+
+export function buildReport(opts: Options): ByteCensus {
+    const userRules = censusRuleDir(opts.userRulesDir);
+    const projectRules = censusRuleDir(opts.projectRulesDir);
+    const claudeMd = censusClaudeMdHierarchy(opts.repoRoot, opts.userHome);
+    const globalProfile = censusGlobalProfile();
+    const skillsCatalog = censusSkillsCatalog(opts.skillsDir);
+    const topRules = topRulesByCost(opts.projectRulesDir, 10);
+    const duplicateScope = censusDuplicateScope(opts.userRulesDir, opts.projectRulesDir);
+
+    const scan = scanTranscripts({ root: opts.root, maxAgeDays: opts.maxAgeDays });
+    const coldStart = computeColdStarts(scan.records);
+    const measuredColdStartMedian = coldStart.legs > 0 ? coldStart.median_first_call_written_or_uncached : null;
+
+    return buildByteCensus({
+        userRules,
+        projectRules,
+        claudeMd,
+        globalProfile,
+        skillsCatalog,
+        topRules,
+        duplicateScope,
+        measuredColdStartMedian,
+        coldStartLegs: coldStart.legs,
+    });
+}
+
+function pct(v: number): string {
+    return `${(v * 100).toFixed(1)}%`;
+}
+
+export function renderText(r: ByteCensus): string {
+    const out: string[] = [];
+    out.push('Preamble byte census (road-to-cache-economy Phase 3, steps 1-2)');
+    out.push('');
+    out.push('Sources:');
+    for (const s of r.sources) {
+        out.push(`  ${s.name}`);
+        out.push(`    files=${s.files} chars=${s.chars} tokens_estimate=${Math.round(s.tokens_estimate)} [${s.provenance}]`);
+        out.push(`    basis: ${s.basis}`);
+    }
+    out.push('');
+    out.push(`measurable_tokens_total (excludes residual): ${Math.round(r.measurable_tokens_total)}`);
+    out.push(`grand_total_tokens (includes residual): ${Math.round(r.grand_total_tokens)}`);
+    if (r.measured_cold_start_median !== null) {
+        out.push(`measured_cold_start_median: ${Math.round(r.measured_cold_start_median)} (${r.cold_start_legs} legs, this window)`);
+        out.push(`measurable_share_of_median: ${pct(r.measurable_share_of_median as number)} (the honest, non-tautological number)`);
+        out.push(
+            `sum check (grand_total vs median, ±10%): ${r.within_tolerance ? 'PASS' : 'FAIL'} — trivially PASS by ` +
+                'construction once the residual bucket exists; see the module doc comment for why.',
+        );
+    } else {
+        out.push('measured_cold_start_median: pending — no subagent legs observed in this window; no residual computed.');
+    }
+    out.push('');
+    out.push('Top-10 project-scope rules by per-spawn token cost:');
+    for (const rule of r.top_rules_by_cost) {
+        out.push(`  ${rule.file}: chars=${rule.chars} tokens_estimate=${Math.round(rule.tokens_estimate)}`);
+    }
+    out.push('');
+    out.push('Duplicate-scope census (C-2):');
+    if (r.duplicate_scope.evaluable) {
+        out.push(
+            `  ${r.duplicate_scope.shared_filenames} shared filename(s), ` +
+                `${Math.round(r.duplicate_scope.duplicate_tokens_per_spawn)} redundant tokens per spawn.`,
+        );
+    } else {
+        out.push(`  not evaluable — ${r.duplicate_scope.reason ?? 'unknown reason'}`);
+    }
+    out.push('');
+    out.push('Modelled duplicate-copy removal (MODELLED, not measured):');
+    if (r.modelled_duplicate_removal.applicable) {
+        out.push(`  modelled_new_median: ${Math.round(r.modelled_duplicate_removal.modelled_new_median as number)}`);
+        out.push(`  modelled_reduction_pct: ${pct(r.modelled_duplicate_removal.modelled_reduction_pct as number)}`);
+    } else {
+        out.push(`  not applicable — ${r.modelled_duplicate_removal.reason ?? 'unknown reason'}`);
+    }
+    out.push('');
+    out.push('The per-spawn payload ceiling (median ≤ 40k / p95 ≤ 50k) is a DOCUMENTED CANDIDATE, gated on C-3');
+    out.push('(preamble reducibility, still pending — no reduction intervention has landed). See');
+    out.push('docs/contracts/load-context-budget-model.md. This report does NOT enforce it.');
+    return out.join('\n');
+}
+
+export function main(argv: string[] = process.argv.slice(2)): number {
+    const opts = parseArgs(argv);
+    const report = buildReport(opts);
+    if (opts.format === 'json') {
+        process.stdout.write(JSON.stringify(report, null, 2) + '\n');
+    } else {
+        process.stdout.write(renderText(report) + '\n');
+    }
+    return 0;
+}
+
+if (process.argv[1] !== undefined) {
+    const invokedUrl = pathToFileURL(path.resolve(process.argv[1])).href;
+    if (import.meta.url === invokedUrl) process.exit(main());
+}

package/src/scripts/probe_skill_registration.ts

@@ -416,7 +416,7 @@ export function run_probe(opts?: {
     // Group by (tool, skill_id). ≥ 2 entries → DUPLICATE.
     const by_key = new Map<string, Registration[]>();
     for (const reg of result.registrations) {
-        const k = `${reg.tool}${reg.skill_id}`;
+        const k = `${reg.tool}\0${reg.skill_id}`;
         const arr = by_key.get(k) ?? [];
         arr.push(reg);
         by_key.set(k, arr);
@@ -426,7 +426,7 @@ export function run_probe(opts?: {
         if (regs.length < 2) {
             continue;
         }
-        const [tool, skill_id] = composite.split('');
+        const [tool, skill_id] = composite.split('\0');
         const key = `${tool}:${skill_id}`;
         result.duplicates.set(key, regs);
         const hashes = new Set<string>();

package/src/scripts/redact_hook_capture.ts

@@ -51,6 +51,19 @@ const _USER_CONTENT_KEYS = new Set<string>([
     'submitted_prompt', 'submittedprompt',
     // Free-form transcript bodies (path stays — content is in another file)
     'transcript', 'transcript_text',
+    // Additive hardening (road-to-global-user-memory Phase 3, 2026-07-30):
+    // the global observation buffer's project-attribution footer
+    // (`user_global_observations.ts`'s `context.project_path` /
+    // `context.project_name` / `seen_in`) is a partial project map by
+    // design (see agents/settings/contexts/global-user-memory-cut.md § the
+    // replacement's own metadata surface) — if any tool call ever passes an
+    // observation object through as an argument, a hook capture of that
+    // call must not preserve these keys the way `_ENVELOPE_KEYS_KEEP`
+    // preserves genuinely structural fields like `cwd`. Reusing this file's
+    // existing key-name match (not a second redactor) mirrors the ADR-103
+    // hidden_unicode precedent: additive, and inputs without these keys are
+    // unaffected.
+    'seen_in', 'project_path', 'project_name',
 ]);
 
 // Keys whose value is a structural / schema marker — keep as-is even when