Release 0.8.0: remediation threads, GitLab commands, and docs

[xeonvs]

Release target

• Toolkit: 0.8.0
• Exact reviewed commit: e2c5b847a3884edeeeb24123412af51b280faf14
• Exact reviewed tree: cec4be33c6fff27de93fded041cde06a798d1ca7
• OCR target: 1.9.10
• Delivery state: implementation, corrective review, package validation, and exact-head hosted validation complete; protected release delivery is ready to continue

Added

• ocr.review-context-policy/v2 with optional remediation_threads; policy v1 remains accepted for existing discussion/reference configurations.
• Private ocr.context-store/v2 remediation records and read-only remediation_thread projections through context_list and context_get.
• Exact whole-reply GitLab lifecycle commands @<live-bot-username> suppress and @<live-bot-username> resolve, with slash-command parity.
• Mode-oriented GitLab examples, an Accepted project-decisions creation/consumption walkthrough, complete environment/default documentation, and navigation indexes.
• CI-configurable OCR_MAX_TOOLS with OCR 1.9.10's default of 30, plus explicit local-only private-artifact retention for DLP diagnosis; validated GitLab MR execution rejects retention.
• Meaningful fault coverage for private results, preflight, GitLab transactions, context/DLP/approval, MCP admission, and provider-neutral contracts.

Changed

• Bot identity and username come from authenticated GitLab GET /user; verified remediation roots require the live bot ID and toolkit marker/fingerprint.
• All MR-derived and adapter/reference text crosses normalization, DLP, and budget admission before private store or model projection.
• Safe non-remediation context remains approval-neutral. Admitted remediation forces comment-only; rejection, mutation, and required degradation cannot enable approval.
• Generic discussion and remediation projections come from one stable twice-read snapshot.
• GitHub Actions cleanup uses bounded UTC-day shards and 14/30/60-day completed-run retention.
• Combined branch-aware coverage is gated at 85%, with four risk groups independently gated at 80%, 80%, 85%, and 85%.
• Toolkit-generated background is classified before model execution by the installed preflight-qualified OCR executable under review --preview; the toolkit does not configure or duplicate OCR thresholds.
• Ordinary success and failure remove toolkit evidence, bootstrap, protected rules, context, action receipts/lock, local DLP diagnostics, and the isolated OCR home. Only closed static pre-execution status may remain for posting handoff.

Removed and migration

• Removed: OCR_GITLAB_BOT_USER_ID; authenticated GET /user is authoritative.
• Removed: OCR_USE_ANTHROPIC; set OCR_LLM_PROTOCOL=anthropic.
• Removed: example-only OCR_RUN_HELPER_TESTS and unsupported documentation-only OCR_LLM_SUPPORTS_FUNCTION_CALLING, OCR_LLM_SUPPORTS_REASONING, and OCR_CONFIG_PATH.
• @bot retest remains out of scope because the CI-only toolkit has no GitLab Note Hook receiver; use GitLab retry UI/API.

OCR compatibility

OCR 1.9.9 — inherited predecessor

OCR 1.9.9 was qualified with toolkit 0.7.1 and remains historical evidence only.

OCR 1.9.10 — exact 0.8.0 target

• The only OCR version accepted by toolkit preflight and pinned by the GitLab example.
• Linux amd64 SHA-256: 359e5bafda1438a47ef389399f4994350e1016371eac1dc17a2c428acb228e6c.
• Retry diagnostics remain private and do not feed toolkit telemetry, DLP, receipts, severity, outcome, or approval.
• OCR itself owns current background warning/rejection limits and independently revalidates the same background during review.

Exact-head validation

• All 13 hosted checks pass on e2c5b847: Python 3.12–3.14, macOS/Linux, build, dependency audit/review, Bandit, Gitleaks, and CodeQL.
• Canonical local gate: 1,185 tests plus 277 subtests; 86% displayed combined branch coverage.
• Risk groups: 83%, 81%, 85%, and 87%.
• Ruff, strict MyPy, Bandit, Towncrier/release contracts, pinned Gitleaks, public-diff privacy review, and git diff --check pass.
• Two explicitly versioned 0.8.0 builds are byte-identical. Twine, archive privacy/content, and clean wheel/sdist installs plus CLI smoke pass on Python 3.12–3.14.
• Correct local package hashes: wheel 6bc46d424b5ca4cb412134607a28e8683aabda70cceb85d3bddca68d44167c05; sdist 225ecf88fddfe23be503de109dfe9e9b8b354c815e22e75c1e29613fa5208cca.

OCR review and explicit waiver

The final direct OCR 1.9.10 review completed all 43 selected files at concurrency 2 and --max-tools 60, with no failed, reused, or waived files and no tool-round exhaustion. Static validation and self-review reduced its candidates to five confirmed bounded corrections, all covered by regressions and the gates above.

The separately planned enriched-context receipt was not produced. On 2026-08-24 the owner explicitly waived that remaining qualification and requested no further qualification run. Therefore this PR does not claim external receipt evidence for actual context_list/context_get use, still-present/evidence-resolved model behavior, or the receipt-level leakage audit. The waiver removes that plan-specific blocker; it does not replace deterministic tests, direct OCR remediation review, exact-head hosted checks, or the stable-release registry/provenance/readback gates.

Closes no issues yet. Coordinates #120, #126, and #127; they remain open through stable external reconciliation.