SameWrite 1.3.1 — fail closed on partial history evidence

[ipeterpetrus]

Patch release candidate for SW-1303-PARTIAL-EVIDENCE. Not 1.4. v1.3.0 is not moved and its
released runtime is unaffected.

This PR has been through two full independent adversarial reviews. The first one blocked it.
That history is kept below rather than tidied away: the point of a release gate is that it sometimes
catches things.

1. The original defect, before and after

                                        status              candidate files   exit
before  PARTIAL history, no accept      CANDIDATE                 1             0
before  PARTIAL history + COMPLETE live CANDIDATE                 2             0
after   PARTIAL history, no accept      PARTIAL_EVIDENCE          0             0
after   PARTIAL history + COMPLETE live PARTIAL_EVIDENCE          0             0

tools/optimize.py read evidence quality from the live sweep only, and the trend finding had no
quality gate at all. A history built from bounded sweeps — carry.py --max-files N, the normal way
to keep a large profile affordable — was indistinguishable from one swept in full.

The rule now lives in one place: a finding may be promoted only when the worst quality among the
evidence actually eligible to support it is COMPLETE, or is PARTIAL and the caller passed
--accept-partial. Worst, not majority. Eligible, not everything.

2. The first review blocked this branch

Head bd7ada5 was reviewed on the complete diff and returned 19 hypotheses: 5 HIGH, 3 MEDIUM.
Every one was reproduced locally before being accepted — none taken on the reviewer's assertion.

	what it was	origin
H1	a record claimed COMPLETE while its own counters recorded 5 unreadable files, 2 oversize lines and a file cap, and was promoted; so was one with sessions=turns=carry_bytes=0	pre-existing
H2	a torn JSON line was skipped silently, so a sweep that lost evidence still reported COMPLETE — no history needed to defeat the new gate	pre-existing
H3	a candidate from accepted-PARTIAL evidence was written with effective_evidence_quality: COMPLETE	introduced here
H4	HOST_BEHAVIOR_SHIFT reported status 30 and strict exit 40 while writing a candidate file	invariant introduced here, broken on an uncovered path
H5	the pre-1.3.1 artifact check was a substring search: prose satisfied it, an unreadable file counted as current, invalid UTF-8 crashed the run	introduced here
M1	candidates_existing turned from ids into prose while the schema version stayed 1	introduced here
M2	the newest INVALID record became the comparability anchor, then vanished, stranding nine complete records	pre-existing
M3	bounded carry took newest-by-mtime, bounded listing took discovery order	pre-existing

3. What closed them

• H1 — derive_quality() reconstructs the producer's rule from carry.accumulate(), the one
  place that knows how a sweep went, and quality_of() returns the worst of what a record claims
  and what it can show. Negative, non-numeric or boolean counters make it INVALID; sessions==0
  is INVALID or EMPTY on the producer's own terms; a schema-2 record showing no counters cannot
  claim completeness and is UNKNOWN.
• H2 — scan_full() counts the lines json.loads rejects, accumulate() aggregates them and
  degrades the sweep to PARTIAL, and the count reaches the history record as malformed_lines.
• H3 — provenance is attached where a finding is created, through sampled() and
  from_ledger(). The positional sampled_count stamping is gone; no finding infers its source
  from where it landed in a list.
• H4 — a population the analyser has declared is not one world joins the single promotion gate.
  A status that refuses promotion now leaves nothing on disk, on every path.
• H5 — artifact_metadata() parses the header block structurally, bounded to 64 KiB and 200
  lines, returning PROVEN_CURRENT, PRE_1_3_1_REVIEW_REQUIRED or UNREADABLE_OR_UNVERIFIABLE.
  OSError and UnicodeDecodeError both fail closed. Nothing is rewritten or deleted.
• M1 — candidates_existing is bare ids again; the review states moved to
  candidates_existing_review_required and candidates_existing_unverifiable.
• M2 — comparable() decides eligibility first and anchors only on eligible records, using the
  derived quality rather than the claimed one.
• M3 — carry.bounded_paths() is the single definition of a bounded sample, used by both the
  carry sweep and the listing scan.

4. Coverage

Two frozen matrices, both written with their expected outcomes before implementation: the
original A–O partial-evidence matrix, and a second one for the five blockers. 29 mutants, each
RED when the fix it guards is reversed and GREEN on the real implementation. A harness self-check
that plants three corruptions in the regression harness and requires the harness itself to fail.

645 assertions across fifteen suites          29/29 mutants
81/81 mandatory adversarial attacks           harness self-check 8/8
privacy: 5 canary classes x 6 sinks, 0 leaks  AI-VOS readiness 27/27
original SW-1303 arms: PARTIAL_EVIDENCE, 0 candidate files
AI-VOS path: 6 bounded sweeps -> PARTIAL_EVIDENCE, 0 files; with acceptance -> analysable

Positive controls are part of the gate, because a fail-closed patch can look excellent by refusing
everything: a clean COMPLETE population still promotes, an accepted PARTIAL population still
promotes, and unrelated invalid evidence still does not poison a valid run.

5. Test debts this created, fixed rather than papered over

Four fixtures claimed COMPLETE without the acquisition counters a real sweep always writes. Each
now states them, with an added assertion locking the new rule so the fixture edit hides nothing.
Three mutation patterns moved with the code and were re-pointed. Two mutants were masked by a fix
firing earlier than the branch they guard, and their oracles now isolate that branch.
emit_candidates() gained two result classes, so its callers were updated — including one in
experiments/ that CI caught and my local sweep had missed, which is why that sweep now runs every
step CI runs, by exit code.

6. Scope

CANONICAL_SKILL_DELTA=0, DEFAULT_MODEL_CONTEXT_DELTA=0 — no file under skills/ or hooks/ is
touched. carry.history()'s append protocol is unchanged and the concurrency suites still pass.
DEFECT-ENV-EMPTY-ANSWER is out of scope: it was attributed to the machine's global agent
instructions, not to SameWrite. The deferred LOW findings from the earlier acceptance review are
untouched except where legacy quality handling was inseparable from correct partial-evidence
semantics.

TOTAL_SAVINGS=NOT_PROVEN and WORLD_BEST_CLAIM=NOT_TESTED are unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_013pnJ85a3DYFU2ZW7CqQKCv