Reuse JSON parses in the complete T16 boundary audit

[thakrarsagar]

The T16 invoice boundary audit repeatedly parses the same restored fixtures after PR #3009's compression improvement. This adds an opt-in, source-bound reader adapter in research/t16-invoice-read-cache; the original audit, frozen inputs and selected frontier remain unchanged.

Each read checks current file bytes. Byte-identical contents reuse a private serialized JSON value, returning a fresh mutable value; changed contents are parsed normally. All 54 mutation cases and every existing boundary-audit predicate still execute.

On one shared Linux host with CPython 3.12.15, four fresh processes in baseline/cached/cached/baseline order took 179.276 / 170.213 / 179.694 / 197.756 seconds. Mean runtime decreased 7.19% (188.516 → 174.954 s); mean peak RSS increased 29.8 MiB. Timing ranges overlap, so this is a local observation, not a portable speed guarantee. Each cached run reused 488 of 628 reads. RESULT.json retains all measurements, exact commands, environment and the measured verifier hash.

Validation:

• All four complete emitted receipts are byte-identical to the unchanged original EXPECTED.json (SHA256 e3c0e3c93fbbc6a63e43502dbb98b9e6556abe6c412e0c142f7a202f0e420541).
• Ten controls cover byte invalidation with preserved metadata, fresh nested values, repeated changed contents/restoration, malformed or missing inputs, source tampering and rejection of -O.
• Registered replay passed: python3.12 -B scripts/research_package.py --package research/t16-invoice-read-cache (outcome=completed).
• Package metadata/source pins, git diff --check, and scripts/check_research_commit.py passed; the publication-byte check covered all 21 package/input files at 8c62e5f52f1d172bb8c95e795449205015bb5143.

Task: replay-runtime, priority revision 6; base 463113d675af7410ac91d53a4fed51ec6494eb9a. Canonical investigation. Submission overlap refresh covered 253 open PRs and 2,017 Discussion records with complete pagination, finding no competing T16 reader-cache work; this is not a novelty certificate.

Credit Tairitsu-Aya's audit/PR #2081, Josie Elliston's PR #1959 handoff, Kumagi's PR #3009 (95568ae83a73ace578e3972a01aef55c889e58a2), and jackmgaller's predecessor copying technique. Prepared by thakrarsagar with substantial OpenAI Codex assistance, including separate code review and test implementation. Apache-2.0.

Limitations: no full selected cold replay, new kappa, supplier admission, all-size theorem or independent host timing replication. The cache retains first-seen contents for this audit's restoration pattern; it is not a general-purpose workload cache. The README records the next independent-host check.