Mister-Meeseeks · Oct 23, 2026 · Oct 23, 2026
diff --git a/research/qq16-portable-integration/README.md b/research/qq16-portable-integration/README.md
@@ -0,0 +1,177 @@
+# Portable replay of the complete QQ16 protected-return fixture
+
+**Verifier engineering; no new saving, supplier admission or kappa.** This
+package makes JoyboyBrian's PR4790 six-case integration replay executable on
+native Windows with one source-bound, in-memory path-rendering correction.
+The full original result is compared, including JSON scalar types. Frozen files,
+archives, original inventory/hash checks and mathematical guards are retained.
+
+The failure was already reported by Mainzu in
+[`selected-stock-rho-ledger`](../selected-stock-rho-ledger/README.md) for PR3385,
+and blocked the complete integration attempt in
+[PR4949](https://github.com/CrocSwap/integer-mult-bounds/pull/4949). Mainzu's local
+shim was not a registered reusable replay. This contribution supplies a narrowly
+bound adapter, fault controls and the complete PR4790 replay; it does not claim
+the path issue, return optimization or earlier ownership audit as new discoveries.
+It answers [the explicit handoff](https://github.com/CrocSwap/integer-mult-bounds/discussions/1166#discussioncomment-19137316).
+
+## Exact failure and correction
+
+On Windows / CPython 3.12.0, the unchanged command
+
+```sh
+python -I -B research/qq16-protected-return-caller/verify.py --case fixed
+```
+
+fails before the first return at retained PR2753 `verify.py:75`:
+`ValueError: no omitted or extra predecessor files`. Its inventory renders
+relative paths with the host's separator, whereas `FROZEN_FILES` uses `/`.
+The complete 50-file inventory includes nested names such as
+`vendor/controller.py`, rendered as `vendor\controller.py` on Windows.
+
+The adapter changes exactly this expression in executable memory:
+
+```diff
+-str(path.relative_to(here/'predecessor'))
++path.relative_to(here/'predecessor').as_posix()
+```
+
+It requires the complete original module SHA256
+`65126e788eeea10e830fc583cf4afe0bfc4d67495cbf8ac21de3507dbe624d56`,
+exactly one matching expression, and the complete transformed-module SHA256
+`d3685f34f88b1e2f5f8aaf5fe55bdbaf7f4bb004af8bcf321a2e9e4cb4e74ec9`.
+No path is dropped, renamed in the manifest, case-folded or stripped. On POSIX,
+the replacement gives the same names for this exact inventory. Controls check
+both path flavours against every retained name.
+
+The actual source chain is PR3087 `inputs/pr3055.tar` ->
+`inputs/pr2899.tar` -> `inputs/pr2809.tar` -> `inputs/pr2753/verify.py`.
+The source hashes for each archive boundary are checked in `verify.py`.
+The original PR2809 loader's two allowed complete historical selected records
+remain the only accepted records. Its compatibility controls still reject
+changed kappa and proof supplements. The original PR4790 temporary projection
+supplies its exact historical selected record; the repository's current selected
+record is unchanged.
+
+## Coverage and evidence
+
+The new launcher authenticates **143 live files**: all 136 original inputs and
+all seven files of the retained PR4790 package. `SOURCES.json` binds their Git
+blobs and SHA256 values at main `cef4b52f2a59009865bf8eabde1be3f159063703`.
+PR4790's published head is `38dd6b59d50d346bb34fea02a4ee1ac8d282ebbf`.
+The retained package's own authentication also runs.
+
+A temporary compiler hook recognizes only the exact PR2753 source image and
+must fire exactly once per worker. It is removed in `finally`. The original
+PR4790 aggregator is executed unchanged; a module-local subprocess facade
+redirects only its six recognized worker commands to this launcher, preserving
+order, `-I`, `-B`, optimization mode, fresh bytecode-cache namespaces and each
+90-second timeout. Its original adapter, fixture, numerical execution, metering,
+source census checks and every comparison remain executable predecessor code.
+
+| Original case | Reproduced boundary |
+|---|---|
+| `fixed`, `adopt` | Two nonconstant generations, 24 actual child calls per arm, equal complete endpoints and exact final inventories; 8,046 record Tapes per arm |
+| `stale-engines` | Original disjointness guard rejects after four completed returns |
+| `stale-external` | Original external-inventory guard rejects before any completed return |
+| `output-alias` | Original exclusive-inventory guard rejects after four completed returns |
+| `service-stop` | Deliberate stop after the fifth return; actual controlled disposal checked before fixture cleanup |
+
+Each successful generation retains the original order: four producer calls,
+six service calls, then two later producer calls. Packets contain 512 bits in
+this reduced fixture. The inherited whole-fixture difference is exactly
+`(reads, writes, moves) = (61488, 24576, 49200)`. This reproduces PR4790's result;
+it is not an additional optimization.
+
+The canonical **complete original receipt**, rather than selected output fields,
+has SHA256 `1c9579c7e34614030922e51697b511db7bea0a3e4c3a652a54bed162e3e19a67`.
+The launcher compares it with retained `RESULT.json` after each full replay.
+
+Twelve new failure controls exercise actual source/receipt boundaries: missing,
+extra and changed predecessor files; changed Git blob; omitted predecessor
+manifest entry, live source and CI input; altered adapter source; boolean to
+integer and integer to float substitutions in the full original receipt;
+omitted service cleanup; and changed saving. The positive authenticator checks
+50 frozen files, four proof files and eleven live inputs before and after these
+faults. The retained PR2809 selected-record controls run separately. On Windows,
+the control also requires the original unadapted authenticator to reject at the
+reported inventory boundary; on POSIX it requires both authenticators to agree.
+
+## Reproduce
+
+Python 3.12 standard library; no solver, Git subprocess or network is needed by
+the direct verifier. From the repository root:
+
+```sh
+python -B scripts/research_package.py --package research/qq16-portable-integration
+python -I -W error -B research/qq16-portable-integration/verify.py --json
+python -I -W error -O -B research/qq16-portable-integration/verify.py --json
+```
+
+Expected marker: `PASS_QQ16_PORTABLE_INTEGRATION`. The registered hook executes
+both modes, each with a 300-second bound. The direct `--controls-only` option
+checks the source/receipt boundary without running the integrated cases.
+
+Local verification on October 23, 2026:
+
+- Metadata/source-pin validation and both registered commands pass on Windows /
+  CPython 3.12.0 (119.465 seconds for the complete registered pair).
+- Both adapted direct modes pass on Ubuntu under WSL2 / CPython 3.12.3
+  (75.192 and 74.863 seconds). The unchanged original full PR4790 replay also
+  passes there. Timings describe this host, not a speedup or performance promise.
+- All three captured platform/mode JSON reports compare equal; the registered
+  Windows commands compare the same complete original receipt.
+- All 143 live input SHA256 values are unchanged after verification.
+- The selected orientation helper returns `PASS_RESEARCH_BASELINE_BINDINGS_ONLY`;
+  this checks source bindings, not the selected mathematical cold replay.
+
+The official exporter can snapshot the committed package and its declared
+inputs. After integrity checking and extraction, the same direct commands run
+without Git. There is no write/regeneration mode and no expected receipt is
+updated by verification. Archive checks, publication-byte validation and hosted
+CI status are reported in the linked PR.
+
+This is a shallow checkout, so the optional full-history publication-ancestry
+preflight was not run. Every declared pin resolves at the exact starting main
+commit; no unmerged foreign Git object is required.
+
+## Scope, coordination and credit
+
+This is a reproducibility extension using the original mathematical checker,
+not a second independent implementation or human/formal proof audit. The six
+cases use reduced geometry. Production O73 execution, a new complete root
+census, full native/all-size admission, stock/precision/compiler proofs and the
+complete D* versus B* cost are not supplied. Finite rebinding control remains an
+unknown signed charge. The inherited source census and whole-root saving
+formula remain credited source evidence. No selected cold replay was run.
+
+Handoff-first assignment, `portable-replay`, priority revision 7. The original
+card and refinement are in `TASK.json`.
+[Canonical claim and findings](https://github.com/CrocSwap/integer-mult-bounds/discussions/4744#discussioncomment-19137557).
+Initial remote overlap covered 315 open PRs and 4,035 Discussion records with
+complete pagination (03:44–03:45 UTC, October 23, 2026). Subsequent local search
+covered all 6,601 tracked top-level notes and found the credited Mainzu precedent.
+Canonical main was refreshed to `3fd7d441b898563fedb7dc45e5f926d937d4dfc8`;
+none of the 143 inputs or governing guidance changed. The scientific source
+pins remain at the starting main commit.
+Arbitrary closed diffs, nested notes and unpublished work are outside that scan.
+
+Credit JoyboyBrian for PR4790's installed composition and full receipt; Chris R /
+Clear1nsight for PR4166/4167/4172/4177's caller, transaction, tariff and cost
+lineage; DavidAriasC for PR2753's authenticator; Mainzu for the earlier Windows
+observation; Tekalign for role exchange; Jose Manuel for physical machinery;
+and all QQ16/native predecessors named in the retained package and archives.
+The PR4949 handoff and finite ownership audit retain Mihir Srivastava's credit.
+New adapter, controls and replay by Mihir Srivastava with substantial OpenAI
+Codex assistance. No subagents or independent human review. Apache-2.0.
+
+## Open next questions
+
+- Price the complete D* versus B* comparison at the actual nonterminal C8 band.
+  Start from PR4790 and current #4744 handoffs, with an explicit finite-control
+  tariff and every unmatched native term. The reproduced return saving alone
+  cannot decide that comparison.
+- Other callers of the same frozen authenticator need their own source closure
+  and full-receipt comparison. Mainzu's PR3385 replay note is one concrete
+  starting point. This package does not install a repository-wide patch; check
+  live overlap before extending it. These questions are structured in TASK.json.