fix(augment): gate the rescue's FTS fallback on query-token coverage (#2092)

[punk-dev-robot]

What

Fixes #2092. The zero-result grep rescue's FTS fallback (_rescue in cli/commands/augment_cmd/search.py) returned the first FTS row pointing at a code page with no quality check, so a multi-token pattern where only one common token matched produced a confident Wiki suggests ... line for an unrelated file (repro in the issue: bluetooth codec negotiation fallback riding on "fallback").

Shape

Token coverage, per @RaghavChamadiya's spec in #2092 (comment) (first comment) — not the BM25 floor I originally proposed: SQLite BM25 is unbounded while PostgreSQL ts_rank is ~0–1, so any constant floor silently kills the fallback on one dialect, and SQLite scores are clamp-postprocessed besides.

• New fixed tunable _RESCUE_FTS_MIN_TERMS = 2, documented next to the other rescue tunables.
• Gate in _rescue's FTS loop: needed = min(_RESCUE_FTS_MIN_TERMS, len(_pattern_terms(pattern))) distinct query tokens must appear in the row's title + target_path + snippet (lowered). Reuses _pattern_terms, the same guard family as the widened path's single-token check.
• Single-token patterns still fire (needed = 1): the zero-result rescue deliberately allows generic names, and the gate must not mute that.
• A below-coverage row is skipped with continue (a later row may cover) and logged at debug level — visible under augment --verbose — so we can watch what the gate is eating, as requested.

Measurements

From my local before/after run on the dotfiles repo (the table in #2092): the weak query covers 1/4 terms → now silent; btrfs space monitoring, containerd image cleanup, hyprland monitor workspace all cover ≥2 terms in the suggested row → still rescue. The unit tests encode the weak repro verbatim.

Tests

TestRescueFtsCoverageGate in tests/unit/cli/test_augment_search.py: weak single-token ride silent, covered row fires, gate scans past a weak first row, single-token pattern still fires, non-code page still skipped. Sabotage-verified: with the gate neutralized (needed = 0), the silent-ride and scan-past tests fail.

[punk-dev-robot]

Checked both bot items:

1. _log_search_firing is untouched — git diff main...HEAD has no hunk containing it (looks like a false positive on the signature-change detection). The two named callers in test_augment_served_reads.py pass.
2. Ran the three importing test files plus this PR's: test_augment_fast_lookup.py, test_augment_search_digest.py, test_augment_served_reads.py, test_augment_search.py — 165 passed.

[punk-dev-robot]

Both warnings were artifacts of a whole-file ruff format sweep: main is not format-clean (CI runs ruff check only, never format --check), so the formatter rewrapped _log_search_firing's def line (2→1 lines, same signature) and a line inside _handle_search_post (complexity 21 is pre-existing, not introduced). fe597fb restores main's formatting on every line this PR doesn't own — the diff is now exactly the #2092 change: the tunable, the gated _rescue_fts_text helper, and its tests. All named callers still pass.

[RaghavChamadiya]

Thanks @punk-dev-robot. The rescue fallback now requires meaningful query-token coverage against the candidate title, target, or snippet instead of applying a backend-specific absolute score. The tests cover the acceptance and rejection boundaries. Approved.