fix(convergence): isolate blind context and audit resumed attempts - #993

Scope

Blind Claude collection refuses advisor availability and missing evidence from any native journaled attempt, including earlier attempts omitted from resumed workflow progress. Recollection now invalidates prior accepted artifacts and saves a failed audit even when all transcripts disappear. Journal completions must match one preceding start, with duplicate/orphan/wrong-key rows refused. Both documented launches disable advisor and use strict empty MCP configuration; current complete provenance keys are registered without replacing history.

• Base commit: 0d822e91c95d0f3cb24a76270daff67840b4da06 (authorized rebase).
• Lane: lane:foundation.
• Owned paths: transcript auditor, Claude collector and tests, blind launch recipe/test, lane provenance registry, evidence manifest and dated decision record.

SOTA sources

• uutils/coreutils 0.10.0: installed sha256sum --strict --check, supported missing-file and checksum controls. Existing co-op mask_packet.py SHA-256 93f72ef95878465f9c22a8bc4d42b6f4f94650434db623e093411914700f6de0 supplies disclosed transport hash mappings; GNU tar 1.35 packages the complete attachments.

• anthropics/claude-code v2.1.296: installed executable SHA-256 24972e3bc859fab2b46ed4c1e51f7d6130f06d3bd550811a114640de3370d0de. Its native v2 journal retains started identities; restoring/restored controls grant no identities. Native advisor availability/removal use toolChange:"add"/"remove" near bytes224081124/224080488. Installed CLI/source verify CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 and --strict-mcp-config without explicit servers.

• agent-lab at e070125dae03b4e44484ccb78d2d65057ad38f40: .claude/workflows/layer-verdict-lane.js, vendored workflow SHA-256 fd77b74945e9d2bd0b26b7822aabcde00e7e650c327da16d852fdbfd5984c0d6. The installed native workflow and vendored bytes were rehashed equal; neither workflow nor prompts were changed.

• native-agent-stack maintained main at9b3a39c43363f32ac61e74c674e4ba069593fe08: landed fix(adjudicate): bind OmniRoute transport and disclosure #992 adjudication source and historical registry rows. Complete current keys are derived by executing adjudicate.adjudication_provenance() on both the PR source and an isolated snapshot merged with that main, using the actual corrected audit hash; recording policy/test define all Claude key fields.

Evidence-class table

Local commands run

python3 -B -m unittest tests.test_verdict_lane_vendoring
rc0;14 tests at both head and merged snapshot

python3 -B -m unittest tests.test_transcript_audit tests.test_claude_lane tests.test_blind_lane_launch tests.test_adjudicate tests.test_verdict_lane_vendoring
rc0;197 tests at the corrected head; zero skips

python3 -B -m unittest tests.test_native_maintenance tests.test_practice_references
rc0;61 tests

python3 scripts/validate.py
rc0 at rebased head:70 components,11472 hashed files,4 profiles,239 receipts

python3 scripts/evidence_manifest.py --check
git diff --check
rc0

Full source/registry preservation and red/green logs, manifest checks, PR-range-only gitleaks and pre-push receipts are retained in the correction packet and summarized in FIX-PUSHED.

Decision record

docs/decisions/2026-10-11-blind-lane-protocol.md.

All historical registrations remain intact. The second correction derives complete current Claude keys under both paths and the current adjudication key via repository APIs, binding auditor SHA-256 753b4656d79105360619dc1ee0929831d3ab589229166bf34d9ce204959740eb. Registry SHA-256 77c31ee062ba0aeda09e051654477e1c8758b48a950bc0102badbf17ce44016d,24407bytes, is re-registered. Re-derive keys at landing if source bytes move again.

The earlier four-module162-test report omitted the required vendoring module; the correction includes it. FULL validate is an integrity/scope check. No new paid run, native-state change, blind key/export change or #995 change was performed. Controlled Claude8 reruns wait for this PR's landing; MCP-instruction, token-e2e advisor and publication identity work remain the separately queued follow-up.

The CC's08:00Z cue authorized one rebase push onto0d822e91, leased onf32f3757. Current keys were freshly derived through the repository's own APIs; both main and old-head history lists remain exact prefixes. The final registry digest remains56fa627d76e316d573e03bc381fe8e117240d1b6cc2f4f5a02d328dfc27418d9/22496bytes, and its evidence-manifest entry is refreshed. On the actual rebased head, vendoring14/14, the five modules193/193 and FULL validation11472 hashed files pass; native merge-tree returns0. All seven untouched file blobs and their aggregate patch-id06d9bd8818a3776db9cbf958a362d54d799cf0b9 match the previous head. Range-diff changes are limited to the registry/manifest reconciliation. The designated micro range is a2e3265...

That prior rebase evidence remains historical. The subsequent micro correction is ONE forward commit on a4bd61b, with unchanged base and no additional rebase unless registry-only conflicts require it. The second micro range is a4bd61b...

Cited captures for the controlled read are retained privately in lanes/convergence-practice-work/cited-captures-20261011/ (resealed MANIFEST.sha256981aeb71fa758f6ab0f60fb7b540c3eb7abcca3e7b9cf47699d9c7c5dcb27740), plus pr993-micro-fix-20261011/ for current source and historical derivation/test receipts. The co-op attaches both, with manifests. Private contents-API download metadata is redacted and its original response hash marked redacted in REDACTION-NOTES.md; pinned source content stays unchanged. Full binary, pinned source/API/CLI captures, native transcript/run/journal evidence and the9b3a39c4 input snapshot authenticate the citations; private corpus content is not published here.

Host evidence

No host receipt or platform-status change.

Checklist

• No credential/account identity values printed or copied; Codex config untouched.
• No new dependency, paid call or client dispatch; native source inspected at the pinned version.
• Historical registry entries, transcripts, journals, old collection receipts, sealed keys and peer worktrees preserved.
• Historical corrections/rebase retained; ONE new forward micro correction on exact a4bd61b lease; CC landing follows the authorized read gates.

Capture-chain correction for the a191a1c micro

The sole Astra P2 was missing review attachments. The completed private pr993-capture-chain-fix-20261011/REVIEW-ATTACHMENTS.tar.gz supplies the full cited-captures-20261011, pr993-micro-fix-20261011 and named historical pr993-review-fix-20261011 packets, including the full pinned executable, both call-site captures and all final before/after/current-key/binding receipts. CITATION-CHECK.sha256 checks 545 attached files with native sha256sum --strict --check; the correction directory's final MANIFEST.sha256 seals the archive and correction receipts. The co-op and CC must attach this complete archive and verify both seals in each extracted reader directory before ACK.

Current shared source seal: 981aeb71fa758f6ab0f60fb7b540c3eb7abcca3e7b9cf47699d9c7c5dcb27740 (449 files), superseding d574beaca0c7692ed6f392adefb635e9748504dc1195ec65b809829f2c255653. Current micro source seal: dc8305cc1d1b02557b1808fa7eca30883832c1ef4370b43c365d1637fa7eb5ac (63 files), superseding f14b4567a3206cf31a7330e840219a2dd1a409a18530a229172a90d2c91c044c. REDACTION-NOTES.md retains the unavailable original response hash 9ea862224026dc640573c28f2a5e8aef2c8f6efc1637153ad390515a41dcdfe3 → redacted hash 909c82810092811938b4771fbf60909cf917276ce798e4b940acbc3ee2389154; authenticated workflow content/Git blob remains unchanged. Historical seals identify earlier bytes and are not substitutes for the current attachments.

Source manifests remain byte-identical; each packet includes its original-to-masked MASKED.tsv and transport PACKET-MANIFEST.sha256. The pinned binary's recovery after the helper's substring exclusion is explicitly recorded. Historical receipts retain their original revision identities. This forward correction updates the decision and attachment contract; native runtime/provenance bytes are unchanged. Checks prove local attachment integrity, with subsequent independent reader ACKs and landing still pending.

Follow-ups

• Qualify Git-dependent modules and archive/history failures in a genuine merged checkout; the reviewed archives established no merge-only functional defect.
• Record the pinned client’s managed-host strict-MCP compatibility boundary through supported native qualification; the retained branch probe is not an end-to-end model run.
• Route the separately queued MCP-instruction, token-e2e advisor and publication-identity work to its research owners; controlled Claude reruns retain the existing landing/capacity gates.

Review disposition follows the CC PR-wave ruling: in-scope reproduced P1/P2 defects or false claims at the head are blockers; other findings stay in this list. P1 safety/privacy/money/order findings always block. Subsequent fixes wait for Claude, GPT and Codex on the same head, or an explicit CC release; open items after the second fix round go to CC adjudication. Candidate changes return to the research queue, and #1009 has queue priority.