fix(propose-takes): insert sentinel rows for zero-proposal pages to avoid re-extraction every cycle

[LittleRandom]

Closes #2106

Problem

propose_takes claims an idempotency contract — "an unchanged page never re-spends LLM tokens" — but this only holds for pages that produce ≥1 take. A page yielding zero proposals never gets a take_proposals row, so its idempotency key is never stored. It is re-sent to the LLM every cycle it falls inside the scan window.

In a steady brain, the majority of pages have no takeable claims (concept pages, stubs, transcripts, reference), so the phase spends ~85-90% of its tokens re-extracting pages that will never produce a take.

Fix

Insert a sentinel row with claim_text = '__no_proposals__' when the extractor returns []. The composite unique index on (source_id, page_slug, content_hash, prompt_version) catches it on future cycles:

• Page unchanged → same content_hash → SELECT finds the sentinel → cache hit → skip LLM call
• Page edited → new content_hash → SELECT misses → re-extract (legitimate)
• Prompt version bumped → new prompt_version → full re-scan (fresh prompt)

Verification

Before fix: 89 pages scanned, 0 cached, 1060 seconds
After fix: 94 pages scanned, 93 cached, 6.8 seconds total cycle

~156x faster. Verified on a production brain with PostgreSQL engine.