feat(catalyst): preserve complete Phase 1 context evidence

[pmanko]

Summary

• supplies every prior user instruction, every eligible verified example, the relevant prior failure, and any experimental guidance without item-count caps, ranking, or silent trimming
• records the exact Hub request, token measurement, request digest, and context-capacity rejection for each writer and reviewer call
• keeps validator findings advisory, requires successful execution before a kept query becomes an example, and never sends result rows
• defines retained history as user instructions plus structured query, example, and failure records; raw model replies are not replayed as trusted context
• removes the remaining silent failure-detail truncation and stale guidance-precedence wording

Dependencies

This is stacked on #76 for the readable PostgreSQL surface. Exact request evidence is supplied by pmanko/med-agent-hub#23.

Validation

• full Catalyst Gateway suite: 348 passed, 3 skipped
• focused context/request-evidence suite: 132 passed
• Ruff format and lint: clean
• JSON schemas and git diff: clean

Boundaries

There is no fixed relation count or metadata whitelist. If the complete physical request exceeds a selected model's measured context capacity, the request is rejected and recorded rather than silently reduced and retried. No database, seed, deployment, or live-environment changes are included.

[pmanko]

@copilot review

Please review this stacked R6 change against the Phase 1 decisions in the PR body. In particular, check that it preserves complete eligible context and exact per-model request evidence without introducing relation-count limits, context-item caps, ranking, silent truncation, or a required Pin interface. Please do not treat historical wire names such as approvedViews or pinned/unpinned as current product policy.

[Copilot]

@copilot review

Please review this stacked R6 change against the Phase 1 decisions in the PR body. In particular, check that it preserves complete eligible context and exact per-model request evidenc...

Reviewed and updated in 75ccb1d. The implementation already preserves complete eligible context and exact per-invocation Hub request evidence without ranking, caps, or silent trimming; I only corrected the lingering test wording that implied a required pin-style product interface.