OCR 1.11.1 evidence integrity and trusted context for v0.8.7

[xeonvs]

Scope

Deliver toolkit 0.8.7 with exact OCR 1.11.1 compatibility, evidence-integrity fixes,
bounded evidence search and coverage, protected same-revision CI outcome evidence, and the
release documentation tracked by #157–#163.

This is the feature PR. Stable publication remains gated by a separate protected PR titled
exactly Release v0.8.7 after this PR merges and its TestPyPI development publication is
verified.

Added

• fixed read-only ocr_toolkit_evidence_search and ocr_toolkit_evidence_coverage tools over
  the existing DLP-admitted evidence store;
• exact count-only action receipt v2 and toolkit receipt v6 reconciliation for
  summary/list/get/search/coverage;
• protected review-context policy v3 with bounded provider-neutral ci_outcome evidence for
  exact-head GitLab checks;
• dynamic stable PyPI version, supported-Python, and Apache-2.0 product badges;
• qualified Pug, Verilog/SystemVerilog, and VHDL selection from OCR 1.11.1.

Fixed and changed

• incomplete base/head evidence can no longer manufacture dependency deltas;
• prior-round OCR findings remain unverified until current code, tests, or admitted evidence
  validates them;
• compatibility probes separately verify below-threshold grouping, threshold-crossing semantic
  grouping, high-churn per-file grouping, and partial budget behavior;
• GitLab CI snapshot hashing is provider-order independent;
• evidence search rejects operators and controls both before and after Unicode normalization;
• the evidence store performs one complete live environment redaction pass per string leaf.

OCR 1.11.1 is the sole accepted runtime. The review effort remains medium, concurrency remains
operator-configurable with the public example default unchanged, and the toolkit does not set a
provider-specific completion-token override. OCR remains authoritative for its current grouping,
background, tool-loop, and token-limit contracts.

Trust boundaries

• no forced tool_choice and no arbitrary repository search through the evidence MCP;
• tool arguments/results, prompts, reasoning, provider payloads, credentials, session state, raw
  forge identities, URLs, logs, and artifacts stay outside public output and receipts;
• CI outcomes, search results, coverage hints, repeated findings, and tool counts do not change
  finding lifecycle, severity, suppression, merge authority, or approval;
• DLP, cleanup, immutable refs, receipt validation, and conservative approval remain fail-closed;
• provider acquisition and posting remain at provider edges; evidence/store/MCP/receipt contracts
  remain provider-neutral.

Validation

• complete project gate: 1,386 tests and 367 subtests; 86.26% branch coverage; risk groups
  85%, 82%, 86%, and 87%;
• Ruff format/lint, strict MyPy, Bandit, Gitleaks, OCR compatibility manifest, Towncrier draft,
  and diff hygiene passed;
• three complete OCR 1.11.1 project-rule reviews covered all 29 selected files; confirmed findings
  were remediated and revalidated;
• the final controlled OCR 1.11.1 review completed 1/1 selected file with no finding, warning,
  failed, reused, or waived work and with publication and cleanup passed;
• that model-driven run selected summary/list/get/search/coverage once each. The primary subtotal
  reconciled exactly to three ocr_toolkit_evidence calls; search and coverage each reconciled to
  one dedicated by-tool call. Context mode was off, so no enriched-context qualification is
  claimed;
• production-composed stdio MCP benchmarking improved from 1.95 seconds to 1.51–1.63 seconds after
  removing the redundant redaction pass, while preserving the same private/read-only contract.

Delivery

Exact reviewed feature head: 410194895f85be37e1801b3b92e959177aabf04f.

Tracks #157, #158, #159, #160, #161, #162, and #163 in milestone v0.8.7. These issues and the
milestone remain open until the stable workflow publishes immutable artifacts and records its
release receipts.