Complete source-bound four-provider finite admission for PR4088

research/mixed-provider-admission/COVERAGE.md

@@ -0,0 +1,55 @@
+# Admission coverage and invalidation
+
+This map describes the exact PR4088 words and literal fill carried through the
+retained PR4085 engine. Paths in the caller column are in the reconstructed
+source or runtime tree. `adapt.py` records exact preimages and hashes for every
+change. A completed discovery run is not strict admission.
+
+| Obligation | Actual caller and input binding | Evidence or rejection control |
+| --- | --- | --- |
+| Four local provider definitions, both orientations, shared helper order | Retained PR4088 `verify.py --controls`; `adapt.extend_source`; native full/canonical word identity | Exact source words, widths and profiles; changed word and omitted provider reject |
+| Complete native source and execution products | `replay.bind_cache`, `native_outputs.authenticate_native`, then inherited cache/generated-identity checks | Exact compiler closure; `native`, `ledgers`, `local-replay`, `wrappers`, `phases`; eleven fixed scientific products per provider |
+| Literal selector and per-call mixed dispatch | Inherited reconstruction and parent lift; `new/rebind.py` calls `mixed_binding.validate_rebind`; provider arguments enter both discovery and strict replay | All 3,240 calls have the exact provider, orientation and program; selector `(109*t) % 324 < 109`; omitted/same-width-wrong-program/coherent-dispatch mutations reject |
+| Width, owner capacity and chronology | Retained PR4088 owner/operator gates; independent parent lift; inherited controller audit; `binding.account` | Each of 1,321,920 rows uses the selected width and strict window; compact prefixes, per-call non-aliasing, capacity and physical owner identity; width 7/8 adversarial fixture |
+| Recomputed idle and complete paid profile | `validate_rebind` compares its complement accounting to `CONSTRUCTION.json` and `PLAN.json`; downstream invoice consumes rebound plan | Mixed idle `{3:22783,5:1625,48:138492}`, total mass 23,431,680; stale idle and changed complete profile reject |
+| Phase and native endpoint evidence | Retained PR4088 `operator_verify.py --controls`; native exact recipe/endpoint checks; inherited local and Lemma-G checks for each provider and orientation | 10,416 all-character identities over 112 ports; exact phase mutant, omitted phase stage and false private-work erasure reject |
+| Full stock/reserve/depth cost | Inherited `extension_blocknorm.py`, `blocknorm_certificate.full_stock_l2`, invoice, assembly and `independent_core`; exact rebound native ABI and per-call profile | 632 arbitrary columns, 91 zero buffers, 8 centers; actual local prefix 95 with reserve 128, active boundary depth 18, actual stock l2 1323 with cap 1330; actual global prefix 1478, charged payload 2048, physical inventory 323,069, physical row coefficient 19,549 below 20,161 |
+| Discovery and frozen certificates | Inherited `replay.py --discover` runs bind, emit, audit, invoice, assembly, freeze, independent | Records actual mixed certificate; discovery alone is not accepted as release |
+| Ordinary strict verification | Inherited `verify.py` and its strict `replay.py`, with four provider directories propagated to both calls | All inherited predicates retained; four-provider control targets: 201 assembly, 43 controller, 12 bit, 8 independent owner and 6 geometry controls |
+| Outer admission | Inherited `outer/check_ledger.py`, `outer/independent_check.py`, top-level `verify_release.py` | Rebound complete paid profile and all ordinary prerequisites; reproducing the headline moment root alone is insufficient |
+| Uniform all-size mathematics | Outside this finite engine | General generation/lifting, conditioning, fixed-tape realization and analytic transfer remain unresolved here |
+
+## Two corrected draft reuse gaps
+
+The separate adversarial context found two failures in early versions of this
+successor, before publication. These are not attributed to PR4088 or PR4165.
+
+First, an existing preflight's successful source reconstruction did not by
+itself authenticate the cached admission driver. The final gate requires the
+exact inherited engine SHA256
+`e69184d235d8790a3c520977dd44dcfb73a99d56989825485f922ff61059c8b1`,
+the exact 676-file source inventory, adapter bindings, three review stages and
+all 95 retained engine/helper/native-cache inputs. A changed engine is rejected
+before the source adaptation can consume it.
+
+Second, a native source hash and successful stage flags did not commit the
+produced operators. The phase mutation changed a warp's left phase modulo 4
+and refreshed recipe references. The source's exact recipe checker rejected
+the resulting operator. The final reuse gate compares every scientific product
+to fixed commitments, including when all internal references are coherently
+updated. It never derives an expected commitment from the supplied cache.
+
+The regression includes the latter mutation, missing-stage and endpoint
+controls. The clean release records the actual native executions and inherited
+controls separately; a regression success is not substituted for those runs.
+
+## Evidence categories
+
+* **Original authors' retained evidence:** unchanged THREE native results and
+  predecessor scientific input packages.
+* **Campaign execution:** FOUR native replays, exact source reconstruction,
+  independent reviews and the extended inherited admission.
+* **Separate AI checks:** an independent event/span accounting method and an
+  adversarial implementation review. These are not independent human review.
+* **Not established here:** formal proof, a maintainer-reviewed new frontier,
+  or the complete integer-multiplication theorem.