feat(analysis): mine native recurring-correction candidates

[U1oSun]

Summary

• Add a read-only native Learning Capture method that screens ordinary normalized Task Episodes for repeated exact repair routes without requiring adapter-supplied learningPatternId labels.
• Emit a bounded, privacy-safe review packet and validate evidence-bound match or abstain decisions before constructing candidates.
• Feed accepted recurring-correction signals through the existing Learning Loop owner while preserving provider/legacy signals and excluding protective Hook, Guardrail, and permission interventions.

Why

• Issue/Story: No linked issue; this implements roadmap item LC-05.
• User or maintainer outcome: repeated repair behavior can become a reviewable learning opportunity from native Task Episode evidence instead of depending entirely on provider-specific labels.

Traceability and Scope

• Spec/ADR: docs/specs/2026-08-01-lc05-native-learning-candidate-mining.md
• Acceptance criteria addressed: AC-1 through AC-10.
• Canonical owners changed: Task Episode normalization, Learning Loop candidate construction, package verification.
• Explicit non-goals: no model invocation, no automatic Harness asset mutation, no intervention scheduling, no effectiveness claim, and no native repeated-rediscovery inference in this first slice.

Change Type

• Feature
• Bug fix
• Tests only
• Documentation/community
• Refactor with no intended behavior change
• Dependency, packaging, or infrastructure

Test and Review Evidence

Check	Result
node --test test/native-learning-candidate-review.test.mjs test/learning-loop-review.test.mjs test/session-episode-contract.test.mjs	37/37 passed
node --test test/doc-link-graph.test.mjs	6/6 passed
npm run pack:verify	Passed; 369 npm entries and 392 runtime ZIP entries
npm test	1053 passed, 4 failed, 6 skipped; all 4 failures are local Windows symlink EPERM cases outside the changed modules
git diff --check	Passed
Independent counterexample review	No remaining P1/P2 findings after two revision rounds

Manual or visual evidence: Not applicable; this change has no UI surface.

Risk and Recovery

• Compatibility and cross-platform impact: provider and historical flat learning signals remain supported; missing booleans retain unavailable/false/true semantics. The implementation uses platform-neutral Node.js APIs.
• Package, plugin, schema, or generated-file impact: two Learning Loop modules are explicitly verified in both the npm package and runtime ZIP. No generated documentation graph changed.
• Rollback or recovery path: revert commit b88a04ca810f9c7add43d9a5bc2aced72a51dfe1.
• Residual risk or unverified boundary: four symlink-dependent full-suite cases cannot run under the current Windows account; native mining is deliberately limited to recurring-correction.

AI Involvement

• Level: Assisted.
• Human review and validation: automated evidence and two independent counterexample-review rounds are complete.

Checklist

• I followed AGENTS.md, CONTRIBUTING.md, and the relevant canonical-owner guidance.
• The change is focused and does not include unrelated local or generated state.
• Tests and documentation match the behavior actually delivered.
• Markdown links were checked when documentation moved or changed.
• Cross-platform behavior was considered for Windows, macOS, and Linux.
• Package/runtime verification was run when shipped files or dependencies changed.
• User-facing or compatibility changes are recorded in CHANGELOG.md.
• I have the right to contribute this work under the repository's MIT License.