plugin: enrich lines say why a count is 0, and an edit gets one line of tests (#1507)

[swapnilpaliwal-sd]

Fixes #1507, #1546, #1604

What was wrong

The PostToolUse enrichment is what agents read after every Read, Grep and Edit, and most of it went unused:

• ← 0 on framework-called methods. A route handler, a scheduled job, an @app.before_request hook or a library override printed ← 0, which reads as dead code. The graph already stores why (entry_points, decorations, unresolved call sites by name), and the hook never read it.
• Grep preview order (java: the Grep caller preview names test callers and hides production callers, ordered by name #1507). Callers were named in display order, so test methods took both name slots and production callers hid behind the count.
• Several declarations of one name (java: the Grep enrichment for an overridden method shows the first two declarations by path, printing two test helpers as one line twice and omitting the base #1546). Declarations were ordered by file path, so two test stubs under adapter-* took both lines (printed identically) and the base, the main-source override and the count were dropped.
• Shell greps (plugin: the Bash hook annotates a search over piped output or a non-indexed file with unrelated declarations matched by substring #1604). A grep after a pipe, or over a log or shell script, was looked up as code. The prefix match used case-blind LIKE, so FAIL named fail and PARSER named parserPresent.
• Edit hook. An edit that changed no declaration printed "this edit changed 0 declaration(s)". A body-only edit listed 10-42 readers that such an edit cannot break. The PostToolUse Edit path also re-reported declarations the session had already seen. Separately, impact was asked with the package-qualified target from changed, which impact refuses, so every Java signature edit said "(impact unavailable)".

The change

• hooks/_graphline.py (new): zero_label() prints entry (<reason>), ? framework (@X), 0 resolved, N by name, ? framework (overrides a library method), ? framework (extends B) or ? framework (@X on Owner), in that order of strength. With no signal it prints 0. Each is one indexed query. body_line() prints the tests that reach a body edit and the command, using command_for from test-impact. distinct_paths() gives the shortest path suffix that tells same-named files apart.
• hooks/enrich.py:
  • Grep previews name callers ORDER BY is_test, display and count test callers separately.
  • Declarations of one name are ordered base first (with ⇣ N override(s)), then production, then path, with (+N more declaration(s)).
  • A shell grep after |, or over non-source files, is silent. Constant-shaped words are skipped, a shell grep matches whole names, and the prefix match uses GLOB.
  • A Read no longer prints a header with no lines under it.
  • Edit: silent when no declaration changed. A declaration already reported this session is dropped. Body-only changes get one line.
• hooks/changes.py: the same one line for body-only changes on the Bash and UserPromptSubmit paths.
• Both hooks ask impact with shown_target.
• graph_sql.impact also returns test_ids, an additive key.

Tests

tests/enrich_lines.py: 28 checks, each silence paired with a near-miss control. 21 fail on 0.1.4, and all controls hold there. Suites: enrich_budget 17/17, hosts, fastpath 4/4, directive 12/12, and run.py: java 119/119, python 71/71, csharp 14/14.

Smoke (corpus copies, fresh index, before vs after)

• Java web app, Read of a controller: 7 of 7 callables printed ←0. After: 3 ←entry (http), 3 ←? framework (@ModelAttribute / @InitBinder) and 1 more entry. Grep previews now name production callers first (<two production callers>, …; 5 in tests), and the base interface comes first with its override count.
• Java body edit of a service method: 897 chars (14 reader rows) became 113 chars: graph: body edit of <Service>.<method>: 3 test(s) reach it; run: mvn test -Dtest=<TestClass>. A comment-only edit is silent before and after.
• Python web app: three request hooks went from ←0 to ←? framework (@app.before_request). pytest | grep -E 'FAILED|ERROR' used to name an unrelated error_... helper, and is now silent. A grep over source is still annotated.

Deep scan deferred: no full-corpus or held-out measurement was run; the smoke numbers above are the only corpus numbers.