fix: harden governed learning storage

[zzhang82]

Summary

Hardens the v0.14 governed learning-candidate boundary after post-release review found schema/static-client, storage, concurrency, and governance edge cases.

Fixes

• Normalize empty optional text fields at MCP and storage boundaries.
• Normalize empty claim_signal filters from static-schema clients.
• Validate SQL identifiers in ensure_column() and tolerate duplicate-column migration races.
• Add SQLite busy_timeout on bridge connections.
• Wrap FTS rebuild in a savepoint and add a concurrency smoke regression.
• Add indexed is_learning_candidate suppression with legacy backfill.
• Block direct promote() on learning-candidate records and forged structured learning-candidate content.
• Bump package/docs to v0.14.1 and add release notes.

Validation

• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv run --extra dev pytest -q tests/test_v0141_hardening.py tests/test_storage.py tests/test_learning_candidates.py -> 59 passed
• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv run --extra dev pytest -q -> 243 passed
• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv run python scripts/check_release_contract.py -> ok true
• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv run python scripts/check_public_surface.py -> ok true
• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv run python scripts/check_onboarding_contract.py -> ok true
• UV_PROJECT_ENVIRONMENT=/tmp/amb-v0141-venv uv pip check -> compatible
• Klaus claim-safety review -> PASS

Boundaries

The FTS concurrency regression is a smoke-level check, not a proof of all multi-process interleavings. The public MCP surface remains unchanged at 10 tools.