2 commits

Author	SHA1	Message	Date
Anton Razzhigaev	1a50e7a2c7	fix: triad-wave batch — Playwright-parity evaluate, starvation-free cursor, supersession replay, basis-gated density ... Dispositions of the multi-model triad + scope wave (grok-4.6, fable-5, gpt-5.6-sol; scope sol-xhigh), severity order: - _evaluate_bounded lost Playwright's function-invocation semantics (unanimous MAJOR, confirmed against the installed driver: UtilityScript eval()s a raw-string expression and INVOKES a function-valued result when isFunction is unset). page.evaluate(_MARKDOWN_JS), the health snapshot and every '() =>' user evaluate serialized the uninvoked function to undefined. The wrapper now evals and invokes exactly as the driver does; the wrapper-shape test pins both halves. - The settled-refresh cursor pinned its byte offset on the earliest SETTLED row of a still-RUNNING task (grok/sol MAJOR): each tick re-read the same 5MB window behind one long-lived parent and permanently starved every later terminal-boundary settlement past the cap — the exact class the cursor exists to heal. The offset now always advances; deferred tasks ride a durable bounded map retried every tick (oldest-evicted at 500, disclosed), pinned by a starvation test. - A superseded-by-revision acceptance run was excluded from the prior-run lookup even on an identical paid identity (sol MAJOR): the resubmission fell through to the wallet's binding_dispatch_already_claimed and recorded a synthetic DEGRADED panel instead of the typed identical_acceptance_refused terminal. Identity-matched superseded runs now replay for free with a replayed_from_superseded disclosure — evidence revision stays stale-detection, never a paid-cycle mint (owner decision 5=A). Moved _prior_acceptance_run to acceptance_dialogue.py (loop.py byte debt is shrink-only). - The trailing-JSON scanner missed a real directive after an unmatched brace/quote in the prose prefix (sol MAJOR with executed repro; fable MINOR): the forced rail then leaked the raw protocol JSON — the original O4 defect resurfacing on one input class. On primary-scan failure the parser now retries from the last few line-start '{' anchors (bounded, string-aware); the inline-after-unbalanced-prose case stays a disclosed prose-degrade, pinned. The scope wave's fence-peel O(n*k) shortfall is also fixed (no slicing per peel, 4-fence cap) and ARCHITECTURE's parser description synced to the final implementation. - Legacy/raw-basis density rows stayed authoritative for their 14-day TTL after an upgrade (sol MAJOR, probe-confirmed): resolve_main_token_density now calibrates ONLY on bounded_proxy rows (review/aggregate resolvers keep all rows — their text-heavy witnesses measure the same on either basis); a poisoned 0.55 image-route witness falls back to the neutral 1.0 instead of under-admitting by 45%. Tests seeding the main resolver updated to the modern producer shape (the observer stamps the basis). - acceptance_patch_dispositions computed the unreviewed_delegated_apply headline AFTER truncation (fable MAJOR): a delegated apply among the 25-cap-omitted oldest rows false-negatived the one fact the attest decision (4=A) exists to surface. Headline now derives from the complete row set; pinned with a mixed-row scenario. - write_patch_verdict treated custody.emit's False return as success (grok/fable/sol): a lost attestation row was invisible — the packet's empty section read as 'nothing recorded'. Both the exception AND the False-return paths now disclose custody_row_write_failed on the verdict artifact, and the row lands on the canonical custody root. - find_child_tasks overlay test now pins CONTENT restoration (cost, result, artifacts), not just the id (grok MINOR). Also: reservation test keeps the read-surface pin with the offline rationale (a cost-equality behavioral pin honestly returns None without a live pricing route — fable m5 disposition); scope-wave (a)-items verified against the ratified record (tools_on_path=11=A, salvage-rail in the O4 verdict, dialogue-history=R2, cycle accounting=O11 seams) — prompt gaps, not drift. Suites: 11775 parallel + 618 serial + 694 web green; ruff -F clean; size-ratchet regenerated (loop.py net-shrinks further). Co-authored-by: Ouroboros <311266734+ouroboros-agent@users.noreply.github.com>	2026-08-31 06:35:17 +00:00
Anton Razzhigaev	9f01912ed4	feat(delegate): attest patch apply/reject decisions into the acceptance packet (D-trace) ... Phase O-F of the harness-health sprint (owner decision 4=A): integrate_delegated_patch applied a child's diff on five mechanical manifest fields with zero review facts anywhere on the path — and none exist to carry: the child is forbidden from reviewing its own diff and no host review of the captured bytes runs. Rather than dressing the child's own claim up as a review (a provenance lie) or gating the apply (a convergence tax the owner rejected), the APPLY itself is attested: - _write_verdict additionally lands a typed delegate_run_patch_verdict custody row (family-prefixed — an unprefixed spelling is invisible to _iter_rows, the known dead-letter class) carrying disposition, applied, bounded reason, patch sha and the pipeline (delegated run_<rid> subjects vs subagent children, classified at the one writer). A failed verdict- artifact write is disclosed on the row instead of dying as a silent "". - delegate_evidence.acceptance_patch_dispositions projects those rows as a bounded host-attested acceptance-packet section (capability_deltas shape: cap 20 + exact omitted count, newest kept) with the honest headline unreviewed_delegated_apply when a delegated patch landed. Absence = no disposition recorded, never "reviewed clean"; an unreadable log is the typed evidence_read_failed marker. - review_evidence wires the section beside capability_deltas/substrate. No gate on apply anywhere. The nanny-wake copy was considered and cut: at wake time no disposition exists yet, and a sometimes-present field recreates the absence-ambiguity this section exists to kill. Co-authored-by: Ouroboros <311266734+ouroboros-agent@users.noreply.github.com>	2026-08-30 20:34:32 +00:00