is_test flag: switch test-map / dead-verdict / dead-code-SQL off path-name detection onto the stored flag; per-language test-attr detection

[jamie8johnson]

Follow-ons declared out of scope by #2249 (the parse-time is_test flag, PR landing 693a6198). The flag now exists as a persisted v33 chunks.is_test column and the rank-time demotion knob reads it; these consumers still use the old path/name-only detection and share the same #[cfg(test)]-in-src/ blind spot the flag fixes:

1. cqs test-map / cqs dead --verdict classify test-vs-production via is_test_chunk (path/name) — switch them to the stored is_test flag so a #[cfg(test)] fn in a src/ file is correctly classified.
2. The dead-code SQL build_test_content_markers sniffs #[test] in chunk content, which is structurally null for cfg-test-in-src (the attribute sits outside the chunk body — chunk content starts at fn). Same gap, same fix as (1): read the stored flag instead. (1) and (2) are one lane.
3. Non-Rust attribute detection: the other 53 languages get only the registry path/name layer — no per-language #[test]-equivalent AST detection (pytest markers, JUnit @Test, Go func TestXxx, etc.). The flag's parse-time computation has per-language extension points left open; populate them where a language has a first-class test marker.

Priority: (1)+(2) are the natural immediate follow-up (they make the flag's benefit reach the test-map/dead surfaces, and they're the same blind spot). (3) is lower — the registry patterns already cover conventional layouts for all languages; per-language AST detection only matters for in-file test markers like Rust's #[cfg(test)], which most languages express via file/dir convention anyway.

[jamie8johnson]

Two more residuals from the #2249 fable review (both LOW, non-blocking — failure mode is under-demotion = pre-change status quo, no regression), extending this issue's "unify consumers onto the stored flag" scope:

4. Split rank-time semantics — chunk_importance still uses the registry sniff. apply_testcode_demote and the test_demote rank-signal now read chunk.is_test, but chunk_importance (src/search/scoring/candidate.rs) still classifies via is_test_chunk because CandidateRow doesn't carry is_test. Net: a #[cfg(test)]-in-src/ chunk is now demoted by the knob but NOT importance-demoted. Unifying needs is_test threaded onto CandidateRow (same class as items 1–2 here).

5. Residual detection gap — cross-file #[cfg(test)] mod x; and inner #![cfg(test)]. A file made test-only by a #[cfg(test)] mod foo; declaration in a parent file, plus the acknowledged inner-attribute (#![cfg(test)] as first child of a mod body) shape, aren't detected by the parse-time walk (it sees the item's own ancestors, not a sibling declaration in another file). Self-corpus instances: src/cli/batch/handlers/dispatch_tests.rs, src/cli/watch/adversarial_socket_tests.rs, src/cli/batch/tasks/stress.rs. Their #[test] fns are still caught by the attribute layer; only non-#[test] helper items read is_test=0. Cheap partial: add %_tests.rs to the Rust test_path_patterns (catches 2 of 3 by filename; stress.rs needs cross-file module resolution, genuinely out of scope). NOTE: this class is distinct from the audit's dominant crowders (#[cfg(test)] fns inside src/parser/chunk.rs etc.), which the AST ancestor-walk DOES catch — so parser/store: schema-carried is_test flag on chunks (cfg(test)/#[test] detection) — rank-time consumers cant see cfg-test-in-src today #2249's core benefit is intact.