fix: preserve ingestion proposal inventory through compaction

[jalehman]

What

Make staged ingestion proposals keep one ordered, unique page inventory across context compaction and validate create/update effects against current source state before any proposal fragment is frozen.

Why

Long-running ingestion agents could forget earlier staged pages after compaction, restage duplicate slugs, or classify an existing page as a create. That left otherwise recoverable ingestion attempts with no actionable path forward.

Changes

• Freeze ordered proposal inventory atomically
• Validate effects against current pages
• Preserve frozen inventory through compaction
• Reject soft-deleted targets before staging
• Keep malformed agent data out of prompts
• Share one inventory validation contract

Changes walkthrough

File	Change
src/core/ingestion-proposal-contract.ts	Canonical inventory types and validation
src/core/ingestion-proposal-context-policy.ts	Safe compaction projection and summaries
src/core/minions/agent-job-proposals.ts	Atomic freeze and live effect checks
src/core/ai/tool-loop-context.ts	Generic typed context-policy hooks
src/core/ai/gateway.ts	Compose proposal context policy
skills/*-ingestion/SKILL.md	Require repeated inventory and repairs
test/	Runtime, compaction, and adversarial regressions

Testing

• bun run typecheck
• AI compaction/OpenAI suites: 37 passed
• Runtime, E2E, and skill suites: 156 passed
• JSONB, newline, diff, PII, export, and route guards
• Autoreview: clean, no actionable findings
• Thermonuclear review: clean after contract extraction
• Full local gate exposed an unchanged extract-atoms-page-discovery intra-file timeout; the exact test passes alone and this branch does not modify that test or extraction implementation. Hosted CI remains required for the exact PR head.

[jalehman]

Hosted CI comparison for exact head f65a599:

• All non-matrix gates passed: verify, serial tests, gitleaks, semgrep, JSONB parity, Tier 1, Tier 2, and both slow suites.
• Current master run 31244771235 at exact base 6e4fb03 fails the same matrix shard set: 4, 5, 7, 9, 10.
• Shards 7, 9, and 10 match master exactly: same tests, assertions/errors, received values, and counts.
• Shard 5 matches master exactly: the same 11 failures, including the same extract-atoms timeout cascade.
• Shard 4 reproduced all five master skill-conformance failures, then exceeded the 15-minute job ceiling before a final suite summary. No new assertion signature appeared, and this PR changes no shard-4 test file.

The repository-wide sharding and state-isolation defects are tracked separately as Lore lore-b74.22. Focused changed-path proof remains 193/193 tests passing, typecheck clean, and both strict reviews clean.