Improve TypeSafe policy synthesis and measurable RSI (0.5)

[vantasnerdan]

Summary

Implements the top four improvements from the requested recent-change review:

1. Explicit instruction-stack audits with exact selected source snapshots, scope/authority uncertainty, and conflict/duplication coaching.
2. Immutable paired trial protocols/results plus TypeSafe review; separate outcomes, safety, cost, heldout/control coverage and honest missingness.
3. Preflight retains capability-exception reviews; oversized creation preserves valid JSON and assessment receipts.
4. Conservative candidate/reducer recommendations: abstention/coverage/quality checks, counterevidence and cross-kind source-family deduplication.

Replaces the existing Host RSI guidance paragraph rather than appending more policy; canonical policy, AGENTS and pinned templates remain unchanged. Full ranked top-10 review and API guide are in docs/rsi-improvement-review.md and docs/rsi-trials-and-instructions.md.

Validation observed

• 337 Node tests and 1,156 Python tests passed.
• 41 Python AST modules; actual LSP document symbols across 34 selected modules.
• Targeted baseline regression replay: 2/10 -> 10/10 expected behaviors (8 defects fixed, 2 controls preserved; NOT general behavioral improvement).
• Two independent scoped reviews; caught/fixed disabled oversize retention and source-ID coercion.
• Real installed Host/CLI and live Jev 1.13.0: audit, plan preflight, candidate evaluation, missingness trial review and before/after synthetic diagnostics.
• Candidate policy deliberately NOT promoted: outcome evidence insufficient. One live audit duplicate finding is disputed and documented, not hidden.
• Packaging/import and diff checks passed; CI must pass before merge.

Limits / boundaries

Typed judgments are advisory, not permission, verified achievement or causal gain. Trial registration orders result recording, not authenticated experiment execution; refs/family labels remain caller reports. No raw credentials or private memory are included.

User explicitly authorized PR, merge/push to main and validation.

[vantasnerdan]

Validation review: all three PR CI jobs passed in https://github.com/vantasnerdan/memory-rsi/actions/runs/35468008689 (plugin, cli, clean-bootstrap). Two independent scoped agent reviews reproduced the baseline defects, reviewed each others patches, and verified fixes for disabled oversized retention and nonstring audit IDs; no remaining reproduced P1/P2 in inspected scope. Local 337 Node / 1156 Python tests passed. Actual installed runtime returned Jev 1.13.0 for audit, preflight, candidate evaluation and missingness-trial review. Canonical policy was deliberately retained: insufficient outcome evidence. Live audit false-positive/intentional-mirror findings and all measurement limits are documented in docs/rsi-improvement-review.md. This is not an authenticated human approval or proof of general policy improvement; direct owner request authorizes merge after validation.

[vantasnerdan]

Post-merge validation complete: main is 0f2534a; remote SHA matched and local main is clean. All three main CI jobs passed: https://github.com/vantasnerdan/memory-rsi/actions/runs/35468082898 . Fresh local main rerun passed 337 Node and 1,156 Python tests. Final private runtime installed Host 0.5.0 / CLI 0.6.0: hashes matched all 26 Host modules and 42 CLI assets. Existing configured TypeSafe tools remain available; no replacement server or canonical-policy promotion. The targeted 2/10→10/10 regression improvement is software mechanism evidence, not a general agent-performance or causal policy-improvement claim.