feat(method): quality-floors — 4 zero-turn input-dialect floors (wm2 miss → mechanical checks)#144
Open
pilotspacex-byte wants to merge 4 commits into
Open
feat(method): quality-floors — 4 zero-turn input-dialect floors (wm2 miss → mechanical checks)#144pilotspacex-byte wants to merge 4 commits into
pilotspacex-byte wants to merge 4 commits into
Conversation
…act's own example formats (quality-floors 1/4) Closes the WV1 wm2 failure CLASS for ~zero turns (benchmark evidence: an arm's own suite stayed green on naive timestamps while the spec's examples were Z-suffixed — the aware/naive crash shipped green at 0.80): - add_engine/constants.py: _DIALECT_CLASSES closed registry (v1: one class, aware-iso-timestamp; full-shape regex — bare dates never match), named in __all__ like _GATE_MODES. - add.py: _dialect_gaps PURE helper (frozen section-3 body vs declared section-4 test files, fail-open) + a print-only tests->build crossing warning naming the class and the re-cross repair (never refuses; exit/state byte-identical) + a dialect_gap WARN in check over ACTIVE build/verify/done tasks (measure-not-block). - 6 red-first pins (test_spec_dialect_floor.py, live-CLI board harness); three trees byte-identical; ENGINE_MD5 7f96609e / PKG 710a009f; SEAMS x17 re-pin. Full tooling suite 3367 green. Tripwire lesson recorded in-task: a bare directory token in section-4 "Tests live in:" sweeps add.py itself into the tamper snapshot — the build's own target must never be declared as a test file (honest redo via TESTS re-cross, attempt 1). Task: spec-dialect-floor (oneshot lane, AI-plan-verify freeze @v1), milestone quality-floors. author: Tin Dang
…lary (quality-floors 3+4) Task dialect-check-and-data-vocab (quality-floors levers 3+4, gate PASS, frozen @v1 via ai-plan-verify). Two more zero-turn floors closing the wm2 input-dialect class from the WV1 benchmark: - TASK.md.tmpl `### Deep checks` gains the DIALECT line ("tests speak the same value formats the spec's examples use") — audit-counted by the existing `shallow` machinery, zero engine code; TASK.fast.md.tmpl §6 gains the condensed twin ("input dialect held"). - GLOSSARY.md.tmpl `## Sensitivity classes` + this project's .add/GLOSSARY.md + sensitivity.md skill guide gain the PROSE rule: datetime, money, or timezone arithmetic ⇒ declare `data` (full lane), wm2 evidence cited. No new sensitivity token minted — _project_sensitivity_domain proven () over a fresh render. - Byte fences held by compression, not rebaseline: TASK.md.tmpl 11391B (≤11400), skill pool/tree back under 51885/145974 via ~183B trimmed from sensitivity.md prose. - 5 sibling pins amended via an explicit TESTS re-cross, as the frozen least-sure flag predicted: 4 Deep-checks fixtures now fill the DIALECT placeholder (guarantee-lints, advisor-review-step, advisor-verdict-audit, verify-record-rollup); fast-lane-skips' task-local empty-git-diff guard re-pinned to its durable invariant (skip machinery stays fast-lane-only). - New red suite test_dialect_vocab_lines.py (6 pins). All twins synced incl. the 4th gitignored dogfood tree. Full suite 3374 passed. - §6 OBSERVE: lever-1 false positive found dogfooding — _dialect_gaps scans §3 raw body incl. the engine's own freeze timestamp; left as an open SPEC delta. refs: quality-floors milestone (task 2/3) · benchmark/results/2026-07-wv1-rep0.md author: Tin Dang
… input-format boundary (quality-floors 2) Task fast-lane-boundary-line (quality-floors lever 2, gate PASS, frozen @v1 via ai-plan-verify). The last of the 4 zero-turn floors closing the WV1 wm2 input-dialect class: - TASK.fast.md.tmpl §1 gains, directly after Accept:, a `Boundary:` line — >=1 format-variant per external input shape the tests must speak (e.g. aware vs naive timestamp), or an explicit "none — no external input". - cmd_freeze refuses a task whose Boundary value is still the bare template placeholder (or empty): `boundary_unfilled`, fired in the shared validate block directly after unflagged_freeze — BOTH freeze paths (human and --ai-plan-verify), validate-then-write (zero bytes land on refusal). A task with NO Boundary: line is grandfathered: legacy fast tasks and the full lane gain no new refusal. Placeholder rule = empty or bare unfenced `<...>`, backtick-exempt (mirrors _section_unfilled). - New red suite test_fast_boundary_line.py (9 pins, live-CLI board: refusal exit codes + byte-identical state on refusal + both paths + grandfathering + §3-fence text never triggers). - 2 test_fast_lane_skips fixtures amended via an explicit TESTS re-cross (they freeze fresh-rendered fast tasks, which now scaffold the placeholder line): shared _fill_boundary helper. - ENGINE_MD5 re-aimed 16cd7cca (prior 7f96609e); SEAMS scope-token-grammar pin x18 (:5337); add.py + fast template synced across all twins incl. the 4th dogfood tree. Full suite 3383 passed. refs: quality-floors milestone (task 3/3) · benchmark/results/2026-07-wv1-rep0.md author: Tin Dang
…hipped) All 4 human-picked input-dialect floors from the WV1 wm2 miss are live and gated PASS: 1. spec-dialect floor (f9d2303) — tests->build dialect_gap warning 2. fast-lane Boundary line (b2483be) — freeze refuses boundary_unfilled 3. §6 dialect check line (bb72452) — Deep-checks DIALECT bullet 4. datetime/money/tz => data (bb72452) — sensitivity vocabulary milestone-done wrote RETRO.md; archive-milestone removed it from active state (files on disk untouched). Balance principle held: every floor is a ~zero-turn mechanical check at an existing seam — the lean lane keeps its cost, the failure class is closed without re-added ceremony. One SPEC delta left open (task 2 gate record): lever 1's _dialect_gaps scans the §3 raw body incl. the engine's own freeze timestamp, a self-referential false positive found dogfooding — candidate future task. refs: quality-floors milestone author: Tin Dang
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
Comment on lines
+89
to
+90
| "Least-sure flag surfaced at freeze: [contract] narrow " | ||
| "shape — cost: one re-freeze.", |
| import unittest | ||
| from pathlib import Path | ||
|
|
||
| import add |
| Run: python3 -m unittest test_fast_boundary_line -v | ||
| """ | ||
| import io | ||
| import json |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
quality-floors — 4 zero-turn input-dialect floors from the WV1 wm2 miss
Converts the WV1 benchmark's wm2 root cause — the lean arm's own red suite spoke naive timestamps, a friendlier dialect than the spec's own
Z-suffixed examples, so an aware/naive datetime crash shipped green at 0.80 fidelity while full-ceremony add-main caught it — into narrow, ~zero-turn mechanical floors. The lean lane keeps its −27–33% cost; the failure class is closed without re-added ceremony.The 4 floors (3 tasks, each gate PASS)
aware-iso-timestampvalue no declared §4 test file speaks;add.py checklists gapped active tasks underdialect_gapf9d2303Boundary:line;freezerefuses a placeholder/empty value (boundary_unfilled, both freeze paths, validate-then-write, absent line grandfathered)b2483be### Deep checks— audit-counted by the existingshallowmachinery, zero engine code — plus a condensed fast-template twinbb72452sensitivity.md×3: datetime/money/timezone arithmetic ⇒ declaredata(full lane), wm2 evidence cited; prose, no new sensitivity tokenbb72452Balance principle (human-agreed): every floor is a mechanical check at a seam that already existed — the only new refusal (floor 2) fires at a gate the fast lane already had.
Engine / pins
ENGINE_MD5→16cd7cca(floor 2cmd_freezeguard) ·ENGINE_PKG_MD5→710a009f(floor 1_DIALECT_CLASSESregistry inconstants.py)scope-token-grammarpin re-aimed x18 →add.py:5337sensitivity.mdproseTests
test_spec_dialect_floor.py(6),test_dialect_vocab_lines.py(6),test_fast_boundary_line.py(9)test_fast_lane_skipsgained a shared_fill_boundaryhelper; the fast-lane empty-git-diff guard was re-pinned from a task-local check to its durable invariant (skip machinery stays fast-lane-only)add.py check: 715 passed, 0 failedOpen follow-up (not fixed here)
One
[SPEC · open]delta recorded at floor 3's gate: floor 1's_dialect_gapsscans the §3 raw body, which includes the engine's ownFreeze mode: …<ISO timestamp>stamp — so this very task's crossing warnedaware-iso-timestampon its own freeze stamp (a self-referential false positive found while dogfooding). Candidate fix: scan only §3 fenced blocks, or strip the Status/Freeze-mode lines. Left as a future task.Known CI non-blockers (pre-existing, not introduced here)
contract_report_unrecordedon 3 older risk-proportional tasks — deliberately not fabricated toReported: yestest_project_scope_locklock-reclaim TOCTOU flake (py3.10) — known-waived, passes on re-run