fix(cycle): cache empty propose_takes scans

[AdityaRajeshGadgil]

Summary

propose_takes used take_proposals as both the operator-facing queue and the per-page idempotency cache. That works only when the extractor produces at least one proposal. The valid result [] left no row behind, so unchanged pages could re-spend extractor tokens every autopilot cycle.

This PR splits those two jobs:

• adds take_proposal_page_scans as the successful extractor-call cache, keyed by (source_id, page_slug, content_hash, prompt_version) and written even for []
• keeps take_proposals as proposal rows only, with dedup moved to (source_id, page_slug, content_hash, prompt_version, claim_text) so same-page multi-claim output is not collapsed
• treats malformed or wrong-shape extractor output as retryable failure instead of caching it as empty success
• threads dryRun into propose_takes so dry runs do not call the LLM or write proposal/cache rows
• adds migration v80 plus fresh Postgres/PGLite schema coverage, including explicit Postgres RLS enablement for the new cache table

Test Coverage

• bun test ./test/propose-takes.test.ts ./test/propose-takes-cache-schema.test.ts
• bun test ./test/schema-bootstrap-coverage.test.ts ./test/propose-takes.test.ts ./test/propose-takes-cache-schema.test.ts
• bun run typecheck
• bun run check:all
• bun run build
• git diff --cached --check

bun run ci:local:diff could not run in this environment because Docker is not installed (docker: command not found). The diff-aware gate reached Docker startup before failing.

Pre-Landing Review

Two independent review passes were run before opening this PR. The first review found two blockers — wrong-shape JSON could still be cached as [], and the new cache table needed explicit RLS coverage — plus an insert-count nit. This version fixes all three:

1. strict production parsing now throws on malformed/wrong-shape output, while literal JSON [] remains a valid cacheable empty result
2. migration v80 and fresh schema SQL enable RLS for take_proposal_page_scans on Postgres, with a PGLite override for the local engine
3. proposal inserts use RETURNING id, so proposals_inserted reflects rows actually inserted after ON CONFLICT DO NOTHING

The second review passed with no blocking findings.