[petalas]

Summary

Measured verifier improvement with a scoped byte-comparison equivalence argument; no new kappa. Defer the two mismatch-only SHA256 digests in the publication preflight until exact byte equality fails.

- compare bytes; compute both diagnostic digests on every file
+ compare bytes; compute the same two digests only on mismatch

The complete byte/history preflight on the unchanged 42-file, 54,975,541-byte Gaussian fixture falls from median 280.56 ms to 246.91 ms, an 11.99% reduction. Ten alternating runs per side have disjoint ranges, 277.11–300.94 ms before and 241.00–256.76 ms after. All 20 complete receipts agree byte-for-byte; zero command errors. The intervals include fresh-interpreter startup, imports, every Git operation, file comparison and source-history check. This is warm-filesystem evidence on CPython 3.12.15/macOS arm64, not a cross-platform timing claim.

Only one production conditional changes. Exact bytes, error text, reader cleanup/final status, inventories, optional modes and ancestry checks retain their contracts. The proof and reproduction are in research/lazy-publication-digests/README.md. This contribution was produced by Nick Petalas. Credit ywz65's comparator, Nick Petalas's earlier history preflight, roger-sato's immutable-object isolation, Douglas Colkitt's source reader/runner and the retained mode/cleanup contributors.

Evidence

• Before: the new no-diagnostic-work test fails on the authenticated original implementation. Untimed full-command profiling attributes 34 ms of 274 ms to SHA256, consistent with removing 84 diagnostic digests.
• After: all 36 focused tests pass normally and optimized, with zero skips, failures or errors. The final registered test bodies take 8.724 s and 8.822 s, both below their 120 s command limits. New controls check zero success-path digests, exact mismatch text and both digests, reader closure and empty-suite rejection. Existing controls exercise drift, missing/normalized/binary files, literal names, inventory, final-reader errors, source history, immutable Git objects and execute bits.
• PASS_REGISTERED_RESEARCH_PACKAGE passes. Committed bytes, declared source history and executable modes pass for 13 files, at head b5a021802c045c2cb9417ad8b2933c6fea71eb39.
• The historical bytecode consumer's three registered commands pass directly: its 16-case source-bound differential, 11 bytecode tests and eight registered-isolation tests. Its generic metadata gate has a pre-existing pin mismatch described below; no predecessor pins or evidence are changed.
• Internal read-only equivalence/evidence review found no flags. Independent human and formal proof-kernel review are unclaimed.

From the worktree root:

/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/lazy-publication-digests
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/lazy-publication-digests/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/lazy-publication-digests --check-source-history --check-modes
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/lazy-publication-digests/benchmark.py --scratch-dir "$RUN_DIR" --rounds 10

For timing, set RUN_DIR to an absolute directory outside the checkout, materialize research/physical-gaussian-core plus all its VERIFY.json inputs, and retain full Git history. The benchmark authenticates the original script at base 721c03822c421b7f7fcbb2598993b947c4d88419, uses the same shared reader on both sides, and writes its baseline and generated receipt only to that directory. The complete generated measurement receipt is retained below.

Limitations

This improves publication checking, not the Gaussian fixture's mathematics. No native/selected cold replay, supplier admission, physical-machine implementation or multiplication theorem is claimed. Small packages can remain dominated by Git overhead. SHA256 implementation failure on equal buffers is outside the equivalence claim. The checkout must remain stable during checking. Linux/Windows performance and native Windows mode behavior were not run. The autonomous local run playbook stopped at this first measured improvement.

Open next questions

Measure a fixed representative publication workload, including small packages, with all complete receipts preserved. This would determine aggregate benefit; choosing that workload and running it require a separate measurement budget.

Related open work

• #5721 changes Windows ownership in the underlying shared Git reader. This change preserves that reader interface and does not measure its Windows behavior.
• #6156 changes scripts/research_package.py to bound working-source comparison memory during complete metadata validation, using the same Gaussian package as a fixture. This PR changes publication byte/history diagnostics. The commands and measured quantities differ; no combined memory or timing benefit is claimed.
• #6161 changes scripts/find_overlap.py with an exact numeric-reference prefilter under the same replay-runtime card. It measures an overlap-search command, not this publication preflight. No common performance number is compared or combined.
• #6162 reuses authenticated committed metadata in the same scripts/check_research_commit.py publication checker. It removes duplicate Git metadata reads; this PR removes mismatch-only hashing. The fixtures and timed boundaries differ, so percentages are not compared or combined.
• The joint publication-ancestry claim in #1166 targets per-pin resolution and reachability operations, which this change preserves.

CI notes

Uncertified by CI at head b5a021802c045c2cb9417ad8b2933c6fea71eb39. In run 40725862797, plan failed and replay (research/lazy-publication-digests) was skipped. The package replay did not run in CI. A maintainer re-run is required after main is fixed.

The onboarding job ran 680 tests and reported FAILED (failures=2, skipped=23). Its only two failures are inherited documentation mismatches on canonical main 721c03822c421b7f7fcbb2598993b947c4d88419, the PR6030 admission committed with [skip ci]. That admission changed the selected record without matching the status bolding and CI navigation:

• test_research_current_status.CurrentStatusTests.test_current_section_matches_selected_record expects standalone **0.003125804** in docs/research/current-status.md; main instead bolds the larger expression 0.003125804 = 781451/250000000.
• test_research_current_status.CurrentCiNavigationTests.test_current_ci_section_matches_selected_record expects make frontier6030-verify; the test reads docs/ci-verification.md, whose current CI section still names Raise conditional kappa to 0.00310045 with independent helper splits #5599 and make frontier5599-verify.

PR #6120 run 40691776027 fails the same two tests. Open PR #6050 targets the documentation fix. The edits to scripts/check_research_commit.py and tests/test_research_commit_bytes.py introduced no additional failure in this onboarding step. Local package replay and preflight evidence remain separate from hosted certification. No code or documentation was pushed after this CI failure.

The dependent historical package research/isolated-descendant-bytecode stops locally at working source drift: scripts/check_research_commit.py. Its live pin expects SHA256 ecf206a82bf1f4e748efb6cc639919582050d173b3537ec12f7265a4f7b55ead; unmodified canonical base already has 63026ec3805f2a3236cc071fc78cf1c54bc6528c43e2f3d82548dc68bf346bce. The new script is 09bd93c5decc662059c6dfd9759ef3a658432db906c9d75b2130f01fb9009b4e. Its verifier consumes the authenticated historical script, and all three registered commands pass directly. The metadata blocker is preserved and disclosed rather than changing pinned predecessor evidence.

Advance CLAIM. Canonical main was refreshed unchanged before publication. This account's open PRs created since the CLAIM were refreshed before DONE. Related shared-verifier work #6156, #6161 and #6162 is listed above. The remaining #6149 through #6163 contributions study different objects and supply no same-instance preflight result compared here.

Merge Danger

Door: two-way. Blast Radius: publication-checker. The production change is one conditional; diagnostic and reader-lifetime regressions are covered. Frozen sources and selected frontier are unchanged.

Generated complete-command benchmark receipt

{
  "base_commit": "721c03822c421b7f7fcbb2598993b947c4d88419",
  "complete_stdout": "PASS_RESEARCH_COMMIT_BYTES {\"commit\": \"721c03822c421b7f7fcbb2598993b947c4d88419\", \"files\": 42, \"package\": \"research/physical-gaussian-core\", \"source_history_commits\": [\"20ecf2891d80b5acf18d980b2164fb44ea7faa90\", \"be4c9c08a44f85e7162e69d220583ba5ed617a81\"]}\n",
  "cpu_count": 18,
  "errors": 0,
  "implementation_sha256": {
    "after": "09bd93c5decc662059c6dfd9759ef3a658432db906c9d75b2130f01fb9009b4e",
    "before": "63026ec3805f2a3236cc071fc78cf1c54bc6528c43e2f3d82548dc68bf346bce"
  },
  "load_after": [
    4.26806640625,
    3.72119140625,
    3.0302734375
  ],
  "load_before": [
    4.37890625,
    3.7333984375,
    3.03076171875
  ],
  "median_fraction_saved": 0.11993171650172396,
  "package": "research/physical-gaussian-core",
  "platform": "macOS-27.0.1-arm64-arm-64bit",
  "python": "3.12.15 (main, Oct 14 2026, 00:55:37) [Clang 23.1.3 ]",
  "samples": [
    {
      "seconds": 0.2824004159774631,
      "side": "before",
      "trial": 1
    },
    {
      "seconds": 0.24783849995583296,
      "side": "after",
      "trial": 1
    },
    {
      "seconds": 0.27711099991574883,
      "side": "before",
      "trial": 2
    },
    {
      "seconds": 0.24742033309303224,
      "side": "after",
      "trial": 2
    },
    {
      "seconds": 0.280220584012568,
      "side": "before",
      "trial": 3
    },
    {
      "seconds": 0.25164479087106884,
      "side": "after",
      "trial": 3
    },
    {
      "seconds": 0.278225583024323,
      "side": "before",
      "trial": 4
    },
    {
      "seconds": 0.24640295817516744,
      "side": "after",
      "trial": 4
    },
    {
      "seconds": 0.27879599994048476,
      "side": "before",
      "trial": 5
    },
    {
      "seconds": 0.24351800000295043,
      "side": "after",
      "trial": 5
    },
    {
      "seconds": 0.28183800005353987,
      "side": "before",
      "trial": 6
    },
    {
      "seconds": 0.24100016593001783,
      "side": "after",
      "trial": 6
    },
    {
      "seconds": 0.28033520793542266,
      "side": "before",
      "trial": 7
    },
    {
      "seconds": 0.2434565417934209,
      "side": "after",
      "trial": 7
    },
    {
      "seconds": 0.281929292017594,
      "side": "before",
      "trial": 8
    },
    {
      "seconds": 0.24607895780354738,
      "side": "after",
      "trial": 8
    },
    {
      "seconds": 0.28078408306464553,
      "side": "before",
      "trial": 9
    },
    {
      "seconds": 0.2511972920037806,
      "side": "after",
      "trial": 9
    },
    {
      "seconds": 0.30093583394773304,
      "side": "before",
      "trial": 10
    },
    {
      "seconds": 0.25676120910793543,
      "side": "after",
      "trial": 10
    }
  ],
  "schema": "lazy-publication-digests-benchmark/v1",
  "summary": {
    "after": {
      "maximum": 0.25676120910793543,
      "median": 0.24691164563409984,
      "minimum": 0.24100016593001783,
      "runs": 10
    },
    "before": {
      "maximum": 0.30093583394773304,
      "median": 0.2805596455000341,
      "minimum": 0.27711099991574883,
      "runs": 10
    }
  },
  "verdict": "faster"
}