corpus: land the reviewed 100-case canary corpus + wired consumption oracle (rebased, single head 0022)

[worldofhacks]

Replays g's PR #51 onto the integration head. PR #51 is CONFLICTING and red on test + secret-scan; none of that was corpus substance.

Two real defects fixed

1. Parallel Alembic head. The branch numbered its consumption migration 0018 with down_revision 0017 — but upstream 0018 is already provider_call_lineage on the same parent. After rebase, migrations/versions/ held two distinct revisions with the identical id 0018. Alembic rejects that outright. Renumbered to 0022 on down_revision 0021; ScriptDirectory.get_heads() now returns exactly ['0022'].

⚠️ PR #50 also claims 0022. Whichever lands second must renumber to 0023 — they cannot both keep the id.

2. The lint gate hid the whole suite. ruff check failed with 4 errors, and that is step 1 of 11 in the test job. So this lane's own 566-line tests/test_consumption_oracle_wiring.py — the test carrying its central claim — had never executed in CI. Fixed (3 over-long comments, 1 unused import). CI will now run it for the first time.

secret-scan was the .gitleaksignore staleness affecting every pre-ec85a45 branch; the rebase clears it. The 507d032 doc conflict is a duplicate of what landed via #48, so the integration head's authoritative 004 medium / 005 low / 006 low win.

Independent verification — I re-derived every claim rather than trusting it

Claim	Verified
pinned manifest digest	07d649d4…52ca252d ✅ matches
case count	100 ✅
categories	PI 20 / DX 18 / TM 18 / SC 15 / DoS 14 / IR 15 = 100 ✅
turn split	79 single + 21 two-turn = 121 physical ✅
all 300 content hashes	✅ via resolve_workload(..., expected_content_hash=…)
oracle bindings	33 synthetic-canary + 26 system-prompt-canary + 14 resource-limit + 23 none + 4 unwired = 100 ✅

Local gates: ruff check clean, ruff format --check 727 files clean, 36 targeted tests pass.

(A first pass hashing raw file bytes reported 100 mismatches — that was my error, not corruption. The loader hashes canonical JSON: sort_keys=True, separators=(",",":"). Recorded here so the next reviewer doesn't repeat it.)

Honest limits — these bound what a 100-case run can conclude

• Only 14 of 100 cases bind to a runtime_wired oracle (the consumption/resource-limit family). Of its three measured dimensions, only some are reachable against a black-box POST /chat.
• 59 cases are canary-anchored — they can only confirm if the canary is genuinely seeded in the live target.
• 4 cases bind to unwired baseline oracles (patient_scope_violation ×2, forbidden_tool_call ×2) and can therefore only ever return INDETERMINATE. That is disclosed, not hidden.
• 23 cases bind to no oracle and are decisive only if the model Judge is enabled — which it is not, and cannot be until three separate preconditions land.
• The digest 07d649… is pinned nowhere in the repo — it must be supplied at authorization time, so an accidental edit to any case file is caught only by an operator-typed value. Worth a follow-up CI assertion.
• Caps conflict: this branch sets all four targets to $50 / 130 / 121 / 0. The capacity lane (769242e, unpushed) narrows the raise to week2 only and its verifier rejects these blobs. A human must pick one authority before either lands. No committed authorization artifact exists for the $50 envelope — authorization-request-week2.json still carries the old $1 / 40 / 60 envelope and corpus id m11-seed-corpus-v1.