[Bug] Workflow findings unvalidated against repository source code; specification self-review insufficient

Bug Description
What the workflow found 19 findings raised, 19 survived independent refutation, 7 Critical. Every one had the same signature — a claim in DOC-312 that nobody re-derived from the code. The internal-coherence self-review the spec had already passed could not have caught any of them. That's the structural point: pointing the finders at the repo instead of the document was the whole difference. The worst three: - §3's counts were produced by a broken method. grep -c counted each file's own helper def line as a claim, missed five call-site-only idioms, and its bare assert arm matched nothing at all — there isn't one ast.Assert node under tests/. The 417 total reconciled during my self-review only because the inflation and the omissions cancelled. I retracted §3 rather than substituting a corrected number. - §6 assigned the interim dual-harness loop to neither gate. As written, the state that section says "must not happen" was the specified default — 12 suites stop running, CI stays green. That is precisely the failure the migration exists to remove. - The §7 claim gate couldn't catch its own target failure. Non-emptiness is a zero/non-zero predicate, blind to partial extraction — test_backend_p4.py would pass it while dropping three claims. Plus: a third invocation point nobody modelled (the CI offline-profile job), a domain fixture requested but never declared (ScopeMismatch, measured), a fixture-ordering rule that reasoned from teardown when setup runs the other way, and -m "not integration" deselecting rather than skipping — so §6's command and its promised outcome were mutually exclusive. DOC-414 — and it is incomplete 22 tasks, writing-plans format, taxonomy-validated, no placeholders. Task 1 is the AST claim extractor; Task 6 deliberately deletes a claim to watch the gate fail. Phase 2 Tasks 20–33 are not written — the twelve per-suite backend conversions. The drafting agent produced the teardown tasks but not the conversions they depend on, and its output was truncated. Marked in the header and at the gap; do not execute past Task 19. I didn't write them from the gap table, because that would repeat the error the review just exposed. Three things you should know I made the same PR scope error twice. #9's body said "4 commits" when it was 5 — the exact mistake I'd fixed on #8 an hour earlier. Corrected, but worth noting it recurred immediately after being pointed out. An unexplained corruption in 603_operations_manual.md. cd .. had become cd ..m — on the pre-demo checklist line, which would break it. I reverted it. I could not attribute it: no workflow agent made an Edit/Write call on that file, and the one write-capable-looking shell command was a read-only grep. Not the workflow, origin unknown. Worth a look, since it's the file your dry-run depends on.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.220
• Feedback ID: 9687b0b2-8f3c-4ba5-908e-d5cadf9a03e2

Errors

[]