Repository navigation
Amend ADR 0005 D4/D5: core Lean and Rocq are §2 recheck legs - #124
Conversation
The owner chose, 2026-11-17, from the selector after #122: "Count it + amend ADR 0005 D4". This records the ruling in the ADR, so the ADR and PROOF-POLICY §2 stop disagreeing on paper. ADR 0005 gains an amendment section recording two rulings: - 2026-11-09: Lean + Mathlib is a generator, never a leg, and its rows stay BRIDGE-PENDING. - 2026-11-17: theory is rechecked in Rocq and Lean with no Mathlib, and the 2026-11-09 ruling does not cover a Lean leg that imports nothing. D4 as amended separates the two Lean roles (core = recheck leg, Mathlib = generator) and adds Rocq as a recheck leg. D5 as amended: the template removal stands, but the repository now has four proof developments. The Rocq leg is instantiated and gated, not a template. PROOF-STATUS, README and CHANGELOG drop the "put to the owner" wording from #122 and point to the amendment. No grade moves: the ledger already followed this ruling. pwg-check PASS; asciidoctor 2.0.20 renders all four files with no warning. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J4QvDf73vaDJwGZ3qMSeH3
|
Fork governance
Scope: the three lanes below are static checks at the named commit. A K9-DEED conformance🟡 advisory · commit
K9 live corpus (L0–L2): 8 finding(s) / scope notes
Scope: K9 components L0–L2 against contract v1.0.0 at the named standards commit. L3 signature verification and routed coordination-dialect files are not certified. Green is limited to this scope. Orange is missing evidence; yellow is grandfathered debt or warnings; grey is an explicit scope exclusion. report.json, badge.json and the validator logs are in the k9-grid artifact on the linked run. Hypatia verdict grid🟢 clean · commit Advisory: 11 row(s), worst 🔴 high (0 critical, 9 high in grid scope) — reported, never grading the change. Unique graded findings across gating and advisory rows: 25 (25 reporter observations; 21 without sufficient identity to corroborate). Row counts remain per reporter; unique totals do not change gating.
Hypatia scan: 13 finding(s)
panic-attack assail: 3 finding(s)
echidnabot: not_applicable — proofs here are gated by their own toolchains (proofs.yml: Idris2, Lean, Agda); no ECHIDNA engine is configured, so this adapter could only ever report not_run
rhodibot (RSR compliance): 11 finding(s)
panicbot: not_applicable — panic-attack already reports directly as its own row in the Hypatia verdict grid; enrolling panicbot would count the same findings twice
sustainabot: not_applicable — reserved fleet slot; no adapter exists (gitbot-fleet README)
glambot (presentation quality): 82 finding(s)
seambot: not_applicable — no seam register is declared (spec/seams/seam-register.json)
finishingbot (release readiness): 127 finding(s)
cipherbot: not_applicable — no cryptography is implemented in this repository; the hashes present are checksum/integrity pins, which cipherbot would report as cryptographic findings
accessibilitybot: not_applicable — no served HTML surface; this is a checker library/CLI with AsciiDoc documentation
Scope: static findings reported by each row's tool at the named commit. The table grades only findings in files this change touches, so the colour answers "does this change introduce anything?", not "what does the whole repository still carry?". Standing findings are counted in each accordion and in report.json; the baseline filter remains the place where acknowledged debt is recorded. The lane state is the worst GATING row; advisory rows are listed and counted but never set it. Purple, red, yellow and blue grade the worst GRADED finding present; orange is evidence missing, malformed or a scanner that did not run; grey is an explicit scope exclusion with its reason. No row is a security certification. report.json, badge.json and the raw findings are in the verdict-grid artifact on the linked run. |
…nt upkeep) (#126) ## Summary This is the upkeep (doctrine 16) that ADR 0005's 2026-11-17 amendment named and left undone. In #124 the amendment says that `just proof-status` and the other recipes D5 touched "still describe the developments they ran on 2026-10-27". This PR brings them in line with the amended D4 role table: Agda is the normative kernel, Rocq and core Lean are §2 recheck legs, Lean + Mathlib is a generator, and Idris2 is the reference. **Stacked PR.** #124 merged into `claude/pwg-names-ladder`, which is still open against `main` as #119. `scripts/check-recheck-legs.sh` exists only on that branch, so this PR targets it. Retarget it to `main` once #119 merges. **One real defect fixed.** Since #122 added the core-Lean leg, `just proof-check-lean4` has failed on every run, whatever the toolchain. Its orphan guard counted `lean4/ratified-order/RatifiedOrder.lean` as a file that no lake package builds. `proof-check-all` failed with it. No proof, ledger entry or grade changed. ## Changes - **`build/just/proofs.just`** - New recipes: `proof-check-rocq`, `proof-check-lean-core` and `proof-check-recheck-controls` run `bash scripts/check-recheck-legs.sh rocq|lean|controls`. `recheck-legs` runs all three, as the `proofs.yml` job of that name does. A missing or wrong-version prover fails, because that is the script's own behaviour; nothing skips. - `proof-check-all` and `verify-all` now depend on `recheck-legs`. **Both now also need `coqc` 8.18.0 and `lean` 4.16.0 on `PATH`.** - `proof-status` lists the four developments by D4 role, with each one's gate, and prints the derived PWG census. It used to grep for a `**Total**` row that `PROOF-STATUS.adoc` no longer has. It no longer counts vendored `.lake/` or `_build/` files. - `proof-check-lean4` orphan guard: it exempts exactly the `LEAN_FILE` that `check-recheck-legs.sh` names, so a stray file beside the leg still fails. If that variable cannot be read, the recipe fails. `proof-check-rocq` gets the same guard for `.v` files. - The "(No proof-check-coq …)" note is replaced by the recipes it asked for. - The one-line descriptions that `just --list` shows now name each recipe's role. Before, several of them showed a stray mid-paragraph line. - **`README.adoc`**: the repository map gains rows for `rocq/` and `lean4/ratified-order/`, plus the D4 roles. The run-the-checks block gains `just recheck-legs`, and its `verify-all` comment is updated. The sentence "Only L1 and L4 carry a statement in all three" now names the three developments it meant (Agda, Lean with Mathlib, Idris2); its blame date of 2026-11-10 predates the recheck legs. Its claim is unchanged. - **`verification/proofs/README.adoc`**: adds the D4 role table, puts the recheck legs and roles in the tree, lists the new recipes, and describes the recheck-leg enforcement. It also fixes a pre-existing editorconfig indent error inside the tree block. - **`CHANGELOG.adoc`**: adds an entry under Unreleased → Fixed. - **Left as history**: `INTEGRATION-REPORT.adoc`, which is marked as a historical report, and ADR 0005's own D5 text, which the amendment supersedes. `EXPLAINME.adoc` has no prover-count prose. ## RSR Quality Checklist ### Required - [x] Tests pass. `just test` is a template stub that tests nothing, so the real checks are listed under Testing. - [x] Code is formatted (`just fmt` is also a stub). editorconfig-checker v3.1.2 is clean on all four files. - [x] Linter is clean. ShellCheck 0.11.0 at `--severity=style` is clean on every tracked `.sh` and on the changed recipe bodies extracted with `just --show`. asciidoctor 2.0.20 with `--failure-level=WARN` is clean on the three `.adoc` files. - [x] No banned language patterns - [x] No `unsafe` blocks without `// SAFETY:` comments - [x] No banned functions (`just proof-scan-dangerous` PASS) - [x] SPDX license headers present (unchanged; aspect test PASS) - [x] No secrets, credentials, or `.env` files included ### As Applicable - [x] Documentation updated for user-facing changes - [x] `CHANGELOG` updated ## Testing Toolchain: Coq 8.18.0 (Ubuntu 24.04 apt, as CI uses), Lean 4.16.0 and ShellCheck 0.11.0 (both checked against the SHA-256 values pinned in `proofs.yml` and `quality.yml`), and just 1.40.0. ``` just recheck-legs -> PASS: Rocq leg (49/49 closed), Lean leg (49/49 axiom-free), 6/6 planted controls caught just proof-check-rocq, coqc absent -> FAIL: coqc not on PATH ... it never skips (exit 1) just proof-check-lean-core, lean absent -> FAIL: lean not on PATH ... it never skips (exit 1) just proof-check-recheck-controls, coqc absent -> FAIL (exit 1) ROCQ_VERSION=8.19.0 just proof-check-rocq -> FAIL: coqc is not version 8.19.0 (exit 1) just proof-check-rocq + stray .v planted -> FAIL: Rocq files no gate checks ... (exit 1) just proof-status -> the role table + census (28: PWG-3 15, PWG-2ii 7, ungraded 6) just proof-scan-dangerous -> PASS bash scripts/pwg-check.sh -> PASS, 29 obligation(s) derived, table and census match bash scripts/test-gate-falsifiability.sh -> 29/29 bash tests/aspect_tests.sh -> PASS=3 FAIL=0 WARN=0 tests/workflows/*_test.sh -> all 5 exit 0 ``` The `proof-check-lean4` orphan guard was tested with a stub `lake` that exits 0, because building the Mathlib packages needs the Mathlib cache, which is not available here: - the **old** recipe on this tree: `FAIL: Lean files belonging to no lake package: …/ratified-order/RatifiedOrder.lean` (the defect). - the new recipe: PASS. With a stray `.lean` planted beside the leg it fails, and with `LEAN_FILE` missing from the script it fails. **Not run locally:** `build-lean`, `build-agda`, and therefore `verify-all` and `proof-check-all` end to end. Agda 2.7.0.1 and the Mathlib cache are not installed here, and those recipe bodies are unchanged. `just --dry-run` confirms the new dependency order: `verify-all` runs build-lean, then build-agda, then recheck-legs, then the honesty gate and the PWG record. Side note, not changed here: just 1.21 (Ubuntu apt) runs recipes from `build/just/*.just` imports in `build/just/`, so every imported recipe, the existing `pwg-check` included, fails under it. just 1.40.0 runs them from the repository root. `.mise.toml` leaves `just` commented out. --- --- _Generated by [Claude Code](https://claude.ai/code/session_01J4QvDgFtn2stHN9poWMAFr)_ Co-authored-by: Claude <noreply@anthropic.com>



Summary
This is a follow-up to #122, which merged into this base. Before #122 merged, the owner chose "Count it + amend ADR 0005 D4" in the selector. That decision settles whether a Lean recheck leg with no Mathlib may count, given the 2026-11-09 ruling "Lean is a generator, not a leg". This PR writes the decision into ADR 0005, so the ADR and PROOF-POLICY §2 stop disagreeing on paper.
No grade moves. The ledger already followed this ruling. The census is unchanged: of 28,
PWG-315,PWG-2ii7,ungraded6.Changes
+foundation-diversetag.just proof-statusand the other D5 recipes still say "three provers" (doctrine 16).RSR Quality Checklist
Required
pwg-check.shPASS, with table and census unchanged--failure-level=WARNon all four filesunsafeblocks without// SAFETY:comments.envfiles includedAs Applicable
docs/decisions/updatedCHANGELOGupdatedTesting
No proof file changed, so no kernel run is needed or claimed.
Generated by Claude Code