100%

Section 04: Rule files — graph-first guidance in 9 domain rule files

Status: In Progress Goal: The 9 domain rule files identified by the original TPR (arc.md, aims-rules.md, typeck.md, types.md, tests.md, impl-hygiene.md, canonicalization.md, patterns.md, compiler.md) each receive a graph-first paragraph. intelligence.md receives a workflow inventory refresh. Other rule files (canon.md, codegen-rules.md, parse.md, llvm.md, etc.) may also benefit from graph-first guidance but were not in the original TPR’s finding scope — they are tracked for future coverage.

Context: Per the 2026-04-14 TPR verification, the original TPR findings (codex-020 through codex-027, gemini-004) identified 9 specific domain rule files that cite reference-repo paths without mentioning the intelligence graph. The graph indexes all 10 reference repos with 24K+ call edges and vector embeddings for semantic similarity — it resolves cross-repo equivalents in sub-second time. A reader following today’s rule-file guidance would open 5 reference repos manually and grep; after §04 lands, the same reader runs one graph query, narrows to 2 concrete files, and then opens those. LEAK:scattered-knowledge is the finding category (codex finding IDs TPR-XX-020 through TPR-XX-025 plus 026/027).

Reference implementations:

  • Ori .claude/rules/intelligence.md:66-75 (Symbol-First Workflow): existing canonical paragraph, adapted here for insertion at other sites
  • TPR finding provenance: codex-020 arc.md:243, codex-021 aims-rules.md:727, codex-022 typeck.md:957, codex-023 types.md:839, codex-024 tests.md:274, codex-025 impl-hygiene.md:313, codex-026 canonicalization.md:37, codex-027 patterns.md:27, codex-005 intelligence.md workflow inventory, gemini-004 general

Depends on: Section 03 (the paragraph cites compose-intel-summary.md as the canonical summary template).

04.1 Draft the graph-first paragraph template

File(s): Template only (no file written in this subsection)

One canonical paragraph shape used in 8 of the 9 target files (§04.2). intelligence.md gets a different treatment (§04.3 refreshes its workflow inventory, not a paragraph insert).

  • Draft the template:

    ## Graph-first, manual second

Before reading the reference-repo paths cited in this rule file, query the
intelligence graph:

- `scripts/intel-query.sh --human similar "<symbol>" --repo rust,swift,koka,lean4 --limit 5`
  — semantic equivalents across reference compilers in sub-second time
- `scripts/intel-query.sh --human callers "<symbol>" --repo ori` — blast radius
  for changes in this domain
- `scripts/intel-query.sh --human file-symbols "<path-fragment>" --repo ori` — the
  module inventory before editing
[+ optional per-file preset bullet, e.g. `ori-arc --limit 5` for ARC-scoped files]

The graph covers Ori plus 10 reference compilers, synced on every commit.
Manual reference-repo reading stays authoritative — but only AFTER the
graph narrows the search. Never cite a graph result without verifying
against the actual source. See `.claude/rules/intelligence.md` for the
canonical when-to-query workflow and subcommand reference, and
`.claude/skills/dual-tpr/compose-intel-summary.md` for the canonical
query protocol used by review-family skills.

    Note: The fenced template above is the FINAL shipped form (post-TPR round 1 fixes: stats removed, footer labels retargeted). The original pre-TPR draft had hardcoded 191K+/505K+ stats and full workflow inventory / summary template labels; those were fixed per TPR findings TPR-04-001-gemini and TPR-04-002-codex.

  • Draft the template (original draft accepted; final form reflects TPR round 1 corrections above).

  • Note per-file tuning — per-file preset mapping decided:

    FilePreset bulletRationale
    arc.mdori-arcARC IR, RC, reuse subsystem
    aims-rules.mdori-arcAIMS lives in ori_arc crate
    typeck.mdori-inferenceType checking + inference
    types.md(none — cross-cutting)Pool + registries span infer/check/pool — no single preset covers all; use file-symbols + similar instead
    tests.md(none — cross-cutting)Matrix testing spans all subsystems
    impl-hygiene.md(none — cross-cutting)Cross-phase invariants span all crates
    canonicalization.mdori-patternsPattern compilation lives here (Maranget)
    patterns.mdori-patternsPattern system core
    compiler.md(none — architecture)Emphasize cross-crate file-symbols instead

  • Subsection close-out (04.1):

    • Template is drafted and per-file preset tuning is decided
    • Update 04.1 status to complete (frontmatter updated)
    • /improve-tooling retrospective 04.1 (inline, no gaps) — the existing .claude/rules/intelligence.md §Subsystem Mapping table maps file-path patterns (e.g., compiler/ori_arc/) to presets; rule files share their names with those subsystems (arc.md → ori-arc), so the preset decision is near-trivial and needs no new helper. No tooling gaps surfaced during §04.1 (pure decision subsection, zero tooling interaction).
    • /sync-claude 04.1 (no-op documented) — template-drafting subsection; no .claude/ artifacts changed. CLAUDE.md already references .claude/rules/intelligence.md §Subsystem Mapping (line 169 Intelligence graph bullet), so no sync needed.
    • Repo hygiene check — diagnostics/repo-hygiene.sh --check detected test.py, a pre-existing scratch file from a parallel session the user elected to preserve (Step 1.95 decision). No §04.1-generated debris; --clean intentionally NOT run.

04.2 Insert into 9 domain rule files

File(s): 9 files to modify with the template from §04.1

Per-file insertion point (verified line numbers from TPR; compiler.md added 2026-04-14 during §02.2 sync-claude retrospective):

  1. .claude/rules/arc.md (268 lines) — after the cross-backend mirrors discussion near line 243
  2. .claude/rules/aims-rules.md (918 lines) — after the reference compilers block near line 727
  3. .claude/rules/typeck.md (1125 lines) — after the rustc/Koka/Gleam/Swift/Zig/Lean citations near line 957
  4. .claude/rules/types.md (933 lines) — after the prior-art table near line 839
  5. .claude/rules/tests.md (286 lines) — after the production compiler test-strategy references near line 274
  6. .claude/rules/impl-hygiene.md (720 lines) — after the aspirational patterns block citing Rust/Zig/Roc near line 313
  7. .claude/rules/canonicalization.md — after the Maranget / DecisionTree consumer citations near line 37
  8. .claude/rules/patterns.md — after the registry + dispatch prior-art references near line 27
  9. .claude/rules/compiler.md (192 lines) — after the ## Source of Truth section’s 10-repo reference list near line 192 (pair with the “Graph reconnaissance — USE FIRST” bullet in CLAUDE.md Compiler Coding Guidelines added in §02.2)

  • For EACH file in the list:

    • Re-read the target line to confirm the insertion point still fits — all 9 verified (minor drift: impl-hygiene.md +1 line from plan’s 313 to actual 314; others exact)
    • Insert the §04.1 template with file-appropriate preset mentioned in the bullet list — per the §04.1 preset table; reference-repo lists also tuned per-file (arc/aims-rules → rust,swift,koka,lean4; typeck → + gleam; types → rust,swift,zig,lean4; tests → rust,go,zig,swift,koka,lean4; impl-hygiene → rust,swift,zig,lean4; canonicalization/patterns → rust,gleam,elm,roc,koka; compiler → rust,swift,zig,lean4)
    • Diff: net change is one ## Graph-first, manual second section added per file; no existing content deleted in the 9 domain files (all purely additive). Post-TPR line-count deltas differ from the initial +199 due to round 1 fixes (stats removal, footer rewording); see commit 6d7bf77a for the authoritative diff.

  • Verify all 9 land: grep -L 'scripts/intel-query.sh' .claude/rules/{arc,aims-rules,typeck,types,tests,impl-hygiene,canonicalization,patterns,compiler}.md returns empty. ✓

  • Subsection close-out (04.2):

    • All 9 inserts verified
    • Update 04.2 status to complete (frontmatter updated)
    • /improve-tooling retrospective 04.2 (inline, negative finding) — 9 inserts felt mechanical but the template shape varies meaningfully per file (preset bullet, reference-repo list tuning, adjacent-section context). An inject-graph-first.py helper would need per-file config (preset, repos, insertion-anchor regex, adjacent-section wording) which approaches the complexity of the Edit calls themselves. Rule files rarely gain new prior-art references — future additions will happen 1-2 at a time over years, not in batches. One-shot work; no helper warranted. Documented negative finding.
    • /sync-claude 04.2 (no-op documented) — Checked .claude/rules/canon.md for line-number citations to the 9 modified files: all cross-references use semantic anchors (typeck.md §PC-2, arc.md §Cross-Phase Invariant Contracts, compiler.md §Architecture, etc.), NEVER line numbers. CLAUDE.md references (.claude/rules/intelligence.md lines 135/169) are also semantic. Zero drift. No sync commit needed.
    • Repo hygiene check — test.py still present (pre-existing scratch file from parallel session, user-preserved per Step 1.95). No §04.2-generated debris.

  • Placement note on compiler.md — plan instruction was “after the Source of Truth section’s 10-repo reference list”, but I placed the graph-first section BEFORE ## Source of Truth for semantic consistency with the other 8 files. Rationale: “graph-first” requires the guidance to PRECEDE the manual-path list, not follow it; the plan’s “pair with ‘Graph reconnaissance — USE FIRST’ bullet in CLAUDE.md” phrasing reinforces this intent. Flagging explicitly for TPR to review.

04.3 Refresh intelligence.md workflow inventory

File(s): .claude/rules/intelligence.md

The current “When to Query” block (lines 18-34) lists 15 workflows. Several review-family skills and commands added since that block was last refreshed are missing (verify-tpr, sync-claude, fix-next-bug, sync-spec, sync-grammar, verify-roadmap — 6 total; /tp-help was already present). After §05 lands, those workflows will query the graph — the intelligence.md inventory should document that.

  • Update the “When to Query” bullet list in .claude/rules/intelligence.md to add 6 workflows (not 7 — /tp-help was already present on line 48; plan’s list overlapped by one):

    • /verify-tpr — placed next to /tpr-review (“TPR triage”)
    • /sync-claude — grouped with other sync commands (“Doc sync”)
    • /fix-next-bug — placed next to /fix-bug (“Bug autopilot”)
    • /sync-spec — grouped with other sync commands (“Spec sync”)
    • /sync-grammar — grouped with other sync commands (“Grammar sync”)
    • /verify-roadmap — placed next to /continue-roadmap (“Roadmap verification”)

  • §04.2 did not surface any rule files that regularly invoke the graph beyond the 9 listed (the inserts ARE their only invocation — guidance, not automation). No additions to §How to Use Results needed.

  • Subsection close-out (04.3):

    • Workflow inventory reflects the post-§05 state (21 bullets covering all known graph-consuming workflows)
    • Update 04.3 status to complete (frontmatter updated)
    • /improve-tooling retrospective 04.3 (inline, FINDING surfaced) — a script scripts/list-intel-consumers.sh that greps .claude/skills/ and .claude/commands/ for scripts/intel-query.sh would detect workflows the graph-query call exists in but aren’t listed in intelligence.md §When to Query. This is legitimate tooling to prevent inventory drift. However: §05 of this same plan (“Missing-trigger skills & commands”) is explicitly the audit/closure pass for this gap — it will add intel-query.sh invocations to the missing consumers, and any automated drift-detector should be built alongside §05’s work (where the before/after set is known). Filing as implementation anchor in §05 rather than adding a new subsection here: §05 should include a completion-checklist item for a list-intel-consumers.sh drift-detector. Documented; see §04.N for cross-reference.
    • /sync-claude 04.3 (no-op documented) — CLAUDE.md line 38 (“Graph-FIRST fact-check”) describes the correctness-verification rule, not the workflow inventory. CLAUDE.md line 169 (“Intelligence graph”) points at .claude/rules/intelligence.md without enumerating workflows. The workflow list is authoritatively in intelligence.md alone; CLAUDE.md needs no refresh.
    • Repo hygiene check — test.py pre-existing, preserved per Step 1.95. No §04.3-generated debris.

04.R Third Party Review Findings

Round 1 (2026-04-14) — dual-source (codex + gemini), merge /tmp/ori-tpr-Zvld2zZO

  • [TPR-04-001-codex][high] .claude/rules/intelligence.md:39 — Remove future-state workflows from the live inventory. Evidence: The six bullets added at intelligence.md:39,42,52,54-56 describe /verify-tpr, /fix-next-bug, /sync-claude, /sync-spec, /sync-grammar, /verify-roadmap as graph-using workflows. .claude/skills/dual-tpr/compose-intel-summary.md:20-23 explicitly lists three of these (/verify-tpr, /sync-claude, /fix-next-bug) as “Planned future consumers (not yet migrated)”; the actual skill files confirm none of the six currently invoke scripts/intel-query.sh. This creates DRIFT between intelligence.md and the SSOT in compose-intel-summary.md. Impact: intelligence.md misrepresents which workflows currently query the graph vs. which are planned post-§05 consumers, hiding §05 scope. Required plan update: Mark each of the 6 new bullets as [planned; tracked by plans/query-intel-adoption §05] or promote them to live only in the same commit that adds the actual graph queries (i.e., as part of §05). Basis: direct_file_inspection. Confidence: high. Resolved: Fixed on 2026-04-14. Added *[planned — tracked by plans/query-intel-adoption §05]* tag to all 6 bullets (/fix-next-bug, /verify-tpr, /verify-roadmap, /sync-claude, /sync-spec, /sync-grammar) in .claude/rules/intelligence.md:39-56. Added a legend paragraph directly under ## When to Query explaining the tag semantics. §05 will promote these to live when the actual graph queries land. DRIFT against compose-intel-summary.md:20-23 closed.

  • [TPR-04-002-codex][low] .claude/rules/arc.md:261 (and 8 peer files) — Retarget the shared footer to the canonical query SSOT. Evidence: All 9 inserted paragraphs use the same closing pointer — “intelligence.md for the full workflow inventory” + “compose-intel-summary.md for the canonical summary template used by review-family skills”. The actual SSOT split is: intelligence.md §When-to-Query + §How-to-Query + §Symbol-First Workflow = canonical query/reference workflow; compose-intel-summary.md = canonical query protocol. Current labels steer readers toward the inventory framing instead of the real SSOT sections. Impact: Mild scattered-knowledge risk — readers looking for authoritative query behavior hit “workflow inventory” framing instead of “canonical protocol”. Required plan update: Revise the §04.1 template footer and reapply to the 9 files, pointing at the right semantic concerns. Basis: direct_file_inspection. Confidence: medium. Resolved: Fixed on 2026-04-14. Footer labels in all 9 files updated: “full workflow inventory” → “canonical when-to-query workflow and subcommand reference”; “canonical summary template used by review-family skills” → “canonical query protocol used by review-family skills”. grep -rl "full workflow inventory|summary template used by review" .claude/rules/ returns empty.

  • [TPR-04-003-codex][low] .claude/rules/compiler.md:198 — Fix stale directional reference after moving the compiler paragraph. Evidence: compiler.md:198 says “blast radius across all 19 workspace crates (see §Crates below)” but after the placement-before-Source-of-Truth move, the ## Crates section is ABOVE at lines 170-180. Codex verified via git show HEAD:.claude/rules/compiler.md that this “below” wording was introduced by the new insertion, not pre-existing. Impact: Small coherence bug that weakens the sound semantic-consistency rationale for the before-§Source-of-Truth placement. Required plan update: Change “see §Crates below” → “see §Crates above” (or remove the directional cue). Basis: fresh_verification. Confidence: high. Resolved: Fixed on 2026-04-14. .claude/rules/compiler.md:198 now reads “see §Crates above”.

  • [TPR-04-001-gemini][medium] .claude/rules/arc.md:257 (and 8 peer files) — Remove hardcoded graph-size stats from 9 rule files. Evidence: The 9 inserted paragraphs each repeat “The graph indexes 191K+ symbols and 505K+ CALLS edges”. VERIFIED: this is duplicated verbatim across all 9 files. VERIFIED: intelligence.md:75 contains an inline comment # Code symbol queries (Ori + reference repos — 32K+ symbols, 24K+ call edges) — stale numbers creating DRIFT within the same file. “synced on every commit” means the numbers are volatile; 10 places-to-update guarantees future drift. Impact: LEAK:scattered-knowledge on a volatile fact — 10 locations must be updated when the graph grows. Visible drift already exists between 191K/505K (9 inserts + CLAUDE.md + compose-intel-summary.md) and 32K/24K (intelligence.md:75 comment). Required plan update: (1) Remove the “191K+ symbols and 505K+ CALLS edges” line from the 9 inserted paragraphs (it’s a redirector, not the SSOT for size). (2) Update intelligence.md:75 stale comment from 32K+/24K+ to 191K+/505K+ to match CLAUDE.md line 169. Basis: direct_file_inspection. Confidence: high. Resolved: Fixed on 2026-04-14. (1) The 9 paragraphs now say “The graph covers Ori plus 10 reference compilers, synced on every commit.” — hardcoded stats removed. grep 191K+\|505K+ .claude/rules/{arc,aims-rules,...,compiler}.md returns empty. (2) .claude/rules/intelligence.md:75 inline comment updated from 32K+/24K+ to 191K+/505K+ matching CLAUDE.md line 169. DRIFT closed. The canonical numeric facts now live in CLAUDE.md line 169 + intelligence.md line 75 + compose-intel-summary.md only — three SSOT locations, not ten.

  • [TPR-04-002-gemini][low] .claude/rules/intelligence.md /verify-roadmap bullet — Name the subcommands consistent with peer bullets. Evidence: The /verify-roadmap bullet says “review agents run intel queries before the rule-file read cycle” without naming subcommands. Peer bullets name specifics (e.g., /sync-claude says “file-symbols on changed crates”; /sync-spec says “callers of affected symbols”). Breaks the established format of the inventory. Impact: Incomplete documentation; inconsistent with peer bullets. Required plan update: Add subcommand names to the /verify-roadmap bullet (e.g., file-symbols/callers/similar to validate architectural claims). Basis: direct_file_inspection. Confidence: medium. Resolved: Fixed on 2026-04-14. .claude/rules/intelligence.md:57 now reads **Roadmap verification** (/verify-roadmap) *[planned — tracked by plans/query-intel-adoption §05]*: \file-symbols`/`callers`/`similar` — review agents validate architectural claims against the graph before the rule-file read cycle`. Format matches peer bullets.

Round 2 (2026-04-14) — dual-source (codex + gemini), merge /tmp/ori-tpr-jmlMqzva

  • [TPR-04-004-codex][medium] plans/query-intel-adoption/section-04-rules-graph-first.md:215 — Correct the §04.N no-deletions completion claim. Evidence: §04.N claimed “205 insertions, 0 deletions” based on the pre-fix diff. The committed change 6d7bf77a includes TPR round 1 fixes (intelligence.md stale-comment replacement + intro paragraph), making the actual aggregate 306/53. The 9 domain-file inserts remain purely additive; the deletions are in intelligence.md (corrections) and plan-file reformatting. Resolved: Fixed on 2026-04-14. Updated §04.N item to accurately describe the per-file breakdown (9 domain files purely additive; intelligence.md had 2 line replacements as corrections).

  • [TPR-04-005-codex][low] .claude/rules/types.md:849 — Align types.md preset with subsystem mapping. Evidence: types.md used ori-inference preset, but its scope (Pool, Idx, TypeFlags, registries) spans ori_types/pool/, ori_types/tag/, ori_types/flags/ — not just src/infer/ and src/check/ which ori-inference maps to. The preset was a weak fit. Resolved: Fixed on 2026-04-14. Replaced ori-inference preset bullet with cross-cutting pattern: file-symbols "ori_types/pool" + similar "TypePool" --repo rust,swift,zig,lean4. Updated §04.1 preset table to show types.md as (none — cross-cutting).

  • [TPR-04-006-gemini][low] .claude/rules/intelligence.md:44 — Missing subcommand names in /fix-next-bug workflow description. Evidence: /fix-next-bug bullet said “blast-radius + similar” without naming callers/callees subcommands. Peer bullets (e.g., /fix-bug Phase 1) use backticked subcommand names. Resolved: Fixed on 2026-04-14. Changed to “callers/callees for blast-radius + similar”.

  • [TPR-04-007-gemini][informational] Multiple files — SSOT compliance and compiler.md placement successfully validated. Resolved: 2026-04-14 — non-actionable informational finding. Confirms objectives (a), (c), (e) are sound.

  • [TPR-04-003-gemini][informational] .claude/rules/compiler.md:189 — Placement of graph-first section before §Source of Truth is logically sound. Resolved: 2026-04-14 — non-actionable informational finding validating the §04.2 placement-note rationale. No fix required.

Round 3 (2026-04-14) — dual-source (codex + gemini), merge /tmp/ori-tpr-DeZ7thTo

  • [TPR-04-008-codex][medium] .claude/rules/intelligence.md:61 — Align /sync-grammar bullet with §05 file-symbols design. Evidence: Bullet said symbols lookup but §05 design (section-05:156-168) specifies file-symbols "compiler/ori_parse/" and file-symbols "compiler/ori_lexer/". DRIFT. Resolved: Fixed on 2026-04-14. Changed to file-symbols on compiler/ori_parse/ and compiler/ori_lexer/.

  • [TPR-04-009-codex][low] section-04:73 — Refresh §04.1 template snapshot to shipped form. Evidence: Template at lines 61-79 still showed pre-TPR wording (hardcoded stats, old footer labels). Shipped paragraphs differ. Resolved: Fixed on 2026-04-14. Updated fenced template to final shipped form + added post-TPR note.

  • [TPR-04-010-codex][low] section-04:129 — Correct stale +199 line-count breakdown. Evidence: Post-TPR line deltas differ from the initial +199 (stats removal changed each paragraph by -1 line). Resolved: Fixed on 2026-04-14. Replaced specific arithmetic with reference to authoritative commit diff.

  • [TPR-04-011-codex][low] section-04:148 — Remove stale /tp-help from missing-workflows list. Evidence: Introduction lists 7 workflows as missing but implementation correctly added only 6 (/tp-help was already present). Resolved: Fixed on 2026-04-14. Removed /tp-help from the missing list, added “(6 total; /tp-help was already present)”.

  • [TPR-04-012-codex][low] section-04:234 — Update §04.N preset reference to match post-round-2 state. Evidence: Completion checklist still listed “types.md → ori-inference” but round 2 changed to cross-cutting. Resolved: Fixed on 2026-04-14. Updated to match current §04.1 table showing types.md as cross-cutting.

  • [TPR-04-013-codex][informational] — Rule-file insertions validated: slim, redirecting, no LEAK:algorithmic-duplication. Resolved: 2026-04-14 — non-actionable. Confirms core implementation is sound; remaining issues are plan-doc drift only.

  • [TPR-04-014-gemini][informational] — types.md preset deviation is architecturally sound. Resolved: 2026-04-14 — non-actionable. Confirms the round 2 cross-cutting change is correct.

  • [TPR-04-015-gemini][informational] — compiler.md placement and SSOT compliance re-validated. Resolved: 2026-04-14 — non-actionable. Third consecutive confirmation.

  • [TPR-04-016-gemini][informational] — Workflow inventory and SSOT compliance verified successfully. Resolved: 2026-04-14 — non-actionable.

Round 4 (2026-04-14) — dual-source (codex + gemini), merge /tmp/ori-tpr-8heRkxq8

  • [TPR-04-017-codex][medium] section-04:3 — Narrow section scope to the 9+1 files actually covered. Evidence: Section title/goal claimed “every cross-repo prior-art reference” but only 9 domain files got paragraphs + 1 (intelligence.md) got a workflow refresh. At least 7 other rule files also cite reference repos without graph-first guidance (canon.md, codegen-rules.md, parse.md, llvm.md, etc.) — beyond the original TPR findings’ scope. Resolved: Fixed on 2026-04-14. Rescoped title, goal, and success criteria to match the 9+1 file set identified by the original TPR. Other rule files noted as future coverage opportunity.

  • [TPR-04-018-codex][low] .claude/rules/canonicalization.md:48 — Add ori_canon/src/patterns/ to Subsystem Mapping. Evidence: canonicalization.md uses ori-patterns preset but the Subsystem Mapping table didn’t cover compiler/ori_canon/src/patterns/. Resolved: Fixed on 2026-04-14. Added compiler/ori_canon/src/patterns/ to the ori-patterns row in .claude/rules/intelligence.md §Subsystem Mapping.

  • [TPR-04-019-codex][low] .claude/skills/dual-tpr/compose-intel-summary.md:20 — Add 3 missing planned consumers to SSOT registry. Evidence: SSOT listed only 3 planned consumers but intelligence.md §04.3 added 6. /sync-spec, /sync-grammar, /verify-roadmap were missing. Resolved: Fixed on 2026-04-14. Added all 3 to the planned-consumer list in compose-intel-summary.md:20.

  • [TPR-04-020-codex][informational] — Core paragraph implementation validated: slim, redirecting, no LEAK. Resolved: 2026-04-14 — non-actionable. Fourth consecutive informational confirming sound implementation.

  • [TPR-04-021-gemini][informational] — Zero actionable issues verified across all 5 review objectives. Resolved: 2026-04-14 — non-actionable.

04.N Completion Checklist

  • All 10 rule files (9 domain + intelligence.md refresh) reference the graph
  • grep -L 'scripts/intel-query.sh' .claude/rules/{arc,aims-rules,typeck,types,tests,impl-hygiene,canonicalization,patterns,compiler,intelligence}.md returns empty
  • Per-file preset tuning verified — matches the §04.1 table (post-TPR round 2 correction: types.md changed from ori-inference to cross-cutting pattern per TPR-04-005-codex). Current mapping: arc.md → ori-arc, aims-rules.md → ori-arc, typeck.md → ori-inference, types.md → cross-cutting (file-symbols "ori_types/pool" + similar "TypePool"), canonicalization.md → ori-patterns, patterns.md → ori-patterns, tests/impl-hygiene/compiler → cross-cutting (no preset).
  • No existing content was deleted in any rule file (diff spot-check on all 10) — the 9 domain-file inserts are purely additive (each +19 to +26 lines, 0 deletions). intelligence.md had 2 line replacements (stale 32K+/24K+ comment → 191K+/505K+, and intro paragraph added under ## When to Query) — these are corrections, not content removals. Aggregate for the 13-file commit 6d7bf77a: 306 insertions(+), 53 deletions(-); the 53 deletions are from plan-file reformatting, not rule-file content removal.
  • ./test-all.sh green — 15,314 passed / 0 failed / 139 skipped / 0 LCFail. LLVM-backend spec-test crash is BUG-04-030 (tracked in fix-BUG-04-039, fix-BUG-04-074); unrelated to §04 docs changes; test-all.sh reports green per its own verdict.
  • python -m scripts.plan_corpus check plans/query-intel-adoption/section-04-rules-graph-first.md returns 0 errors (ran silent = success)
  • Plan sync:
    • Section frontmatter status → complete, reviewed → true
    • 00-overview.md Quick Reference and mission criterion updated
    • index.md updated
  • /tpr-review passed — 5 rounds (4 finding-fixing + 1 clean), 16 actionable findings resolved, both reviewers confirmed zero actionable issues on round 5. Codex 355s/101 events, Gemini 425s/64 events. ASYMMETRY: MODERATE (accepted — walltime comparable). Merge at /tmp/ori-tpr-VJS1GsnR/merged.json.
  • /impl-hygiene-review passed — doc-only scope (no compiler code changes). SSOT/LEAK:algorithmic-duplication concern covered exhaustively by TPR rounds 1-5: codex confirmed across all 5 rounds that the 9 paragraphs are “user-facing prose tuned per-file” that point at the SSOT rather than duplicating it (informational findings TPR-04-003-codex, TPR-04-006-codex, TPR-04-013-codex, TPR-04-020-codex validated this explicitly). No phase boundaries, naming, or dispatch logic involved in doc-only changes.
  • /improve-tooling section-close sweep — Retrospective: §04.1 (template drafting, no tooling interaction), §04.2 (9 inserts, no tooling gaps — one-off work per §04.2 close-out), §04.3 (workflow inventory, surfaced list-intel-consumers.sh idea deferred to §05 per §04.3 close-out). Section-close sweep: the 4 TPR fix rounds surfaced a CASCADE PATTERN (each doc-fix creates plan-doc drift) — this is inherent to doc-heavy TPR and doesn’t have a tooling fix (the plan file IS the drift source). No cross-subsection tooling gaps beyond the deferred §05 consumer-drift-detector. No new tooling warranted.
  • /sync-claude section-close doc sync — CLAUDE.md line 38 (“Graph-FIRST fact-check”), line 135 (“Graph reconnaissance — USE FIRST”), and line 169 (“Intelligence graph”) are semantic references that describe policy, not specific file content. The §04 changes added content TO the files those lines point at — no CLAUDE.md update needed. .claude/rules/canon.md uses semantic anchors (no line numbers) for rule-file cross-references. No drift detected.
  • diagnostics/repo-hygiene.sh --check — test.py pre-existing (preserved per Step 1.95). No §04-generated debris.

Exit Criteria: All 9 target rule files contain a graph-first reference, with per-file subsystem-preset tuning. intelligence.md workflow inventory reflects the full set of post-§05 consumers. ./test-all.sh green.