memory/kb: widen FTS5 snippet window 24 → 60 + nudge agent to follow up with fs_read

[FMXExpress]

Summary

Addresses the actionable finding from PR #308's LOCOMO bench results:

FTS5-bounded snippet window often clips the actual answer line (snippet R@10 ≈ 0.75). That's an actionable finding — either widen snippet windows in PasClaw.Memory.Index or train the agent to follow up with fs_read on retrieved citations more aggressively.

Does both — the two fixes are complementary, not alternatives.

Bench delta

Re-running the PR #308 harness against the bundled alice_synthetic persona with this branch applied:

metric	before (width=24)	after (width=60)
snippet R@10	0.75	1.0
snippet R@1	0.375	0.875
doc R@10	1.0	1.0 (unchanged)

Doc-level recall was already optimal — the layer found the right file; it just wasn't showing the right line in the snippet. Widening the window closes that gap on this synthetic. The remaining snippet R@1 = 0.875 (1 out of 8 missed) is a BM25 ranking artifact, not a snippet-width issue — leaving for a follow-up that tunes the hybrid ranker.

Changes

1. PasClaw.Memory.Index — lift the magic number, widen to 60

Adds FTS5_SNIPPET_TOKENS = 60 as an interface-section const so PasClaw.KB.Index can reuse it (memory_search and kb_search should agree on snippet width — they share a tokenizer and they share the bench). 64 is FTS5's hard ceiling per call; 60 leaves a sliver of slack for the «...» highlight markup. Both Search() overloads (FPC sqldb and Delphi FireDAC) now interpolate IntToStr(FTS5_SNIPPET_TOKENS).

2. PasClaw.KB.Index — reuse the same const

Two snippet(kb_fts, 2, ...) queries now reference FTS5_SNIPPET_TOKENS (picked up from the existing uses PasClaw.Memory.Index). The uses comment is updated so the cross-unit dependency is grep-able.

3. PasClaw.Agent.Prompt.BuildRulesSection — nudge for citation follow-up

Rule #5 (the memory rule) now explicitly tells the model that memory_search / kb_search return bounded snippets, and that if the cited file is right but the answer line might fall just outside the window, follow up with fs_read (or kb_get for the KB) on the cited path. Frames it as: a snippet showing the right file but not quite the right line is a hit, not a miss — i.e. drive the agent toward iteration on citations instead of giving up.

Why both fixes

• Wider window lifts the common case — short answer lines that used to fall just outside a 24-token window now show up in the first hit. Cheap and zero-token cost to the model.
• Prompt nudge handles long-context documents where even 60 tokens can't centre on the right line (e.g. a multi-paragraph note where the relevant fact is several sentences from the search term that anchored the hit). The model now knows to read the underlying file instead of treating the snippet as the final answer.

Test plan

• make clean — no new warnings/errors introduced
• make test-kb-index passes
• make test-agents-md passes
• Re-ran PR bench: LOCOMO-shaped memory-retrieval harness for memory_search #308's harness — numbers above
• Real LOCOMO run (gated on PR bench: LOCOMO-shaped memory-retrieval harness for memory_search #308 + a shape adapter — separate follow-up)

---

Generated by Claude Code