plugin: answer at once from the last graph with stale rows marked; watch what each parser reads (#1594, #1595)

[swapnilpaliwal-sd]

Fixes #1594
Fixes #1595

What was wrong

• python/java/csharp: the first query after an edit answers from the pre-edit graph, because the 10 s wait is shorter than every measured rebuild #1595: a query after an edit waited a fixed 10 s for the background refresh. Rebuilds took longer than that on every project measured, so the wait was spent and the answer came from the pre-edit graph anyway, with one banner line and no way to tell which rows were affected.
• python/java/csharp: an edit under a directory named out, build, target or coverage never makes the graph stale #1594: the refresher's file table pruned one shared list of directory names (out, build, target, coverage, ...) in every language. The Python and C# parsers read those directories (Java reads coverage), so an edit there never made the graph stale and index said "up to date".

The change (plugin only; no engine or parser change)

• ax_fresh.py query runs context, path and impact stale-while-revalidate. The answer comes from the last good graph at once. Every row that lies in a file edited, added or removed since the graph was built ends with (may be out of date); rows from untouched files carry nothing. --json rows get "stale": true plus a top-level freshness object. MCP returns the marked text.
• It waits (up to AXIOMCODE_FRESH_WAIT, now 30 s) only when the answer or the name asked about touches an edited file, the rebuild is not compiling rules, and the last build of this repository (build-seconds, now written by axiomcode-build) says the refresh will finish within the budget. Otherwise it answers immediately. The never-block rule from Query latency: do once what every query repeated; first build reports its stage instead of blocking #1576 and the safe swaps from Refresh after edit: one engine lookup, no silent fallback, safe swaps, .gitignore honoured #1565 are unchanged.
• --fresh (CLI) and fresh=true (MCP context, path, impact) wait for the refresh, print progress every 10 s, and answer from the new graph. SKILL.md tells agents to use it before a delete or a rename.
• The file table prunes, per language, exactly what that language's parser skips (SKIP in ax_fresh.py, taken from the parser's own constants). tests/freshness.py reads those constants back from parser/src, so the two lists cannot drift apart unnoticed.

Tests

• New tests/freshness.py, 24/24 (no engine). Covers: prune lists equal to the parser's, per language; marks in text and --json, with near-miss controls (a same-named file in another directory, quoted code lines, next: lines, an edit the answer never names); the wait decision in each direction, including one control per case where it must not wait; --fresh; MCP fresh.
• tests/refresh.py gets two new checks (marked rows, --fresh). Results: python, java and csharp all pass. tests/latency.py passes (27). tests/mcp.py passes.
• tests/refresh_races.py has 2 failures here ("query mid-build", "no rules"). The same 2 checks fail on unmodified 0.1.4 at a3f0fa3 (the commit this branch is on) on this machine, so they are not caused by this change.

Repros (before = 0.1.4 at f71faef, one merge behind a3f0fa3 that only changes the JavaScript engine; after = this branch)

• The pruned-dir repro (python, java, csharp): FAIL before, PASS after in all three.
• The first-query repro (a 2-file Python project): PASS after in 2 of 2 runs, answered in 4-5 s after a short wait. Before, it failed 3 of 3 earlier runs but passed 2 of 2 later runs. Rebuilds of that tiny project now finish within the old 10 s, so this repro no longer separates the two versions.
• Smoke on a 78-file Python CLI project, adding a caller in a second file:
  • Before: the query waited 10 s and answered without the new caller.
  • After: it waited 9 s (the last build took 7 s) and listed the new caller.
  • With the wait disabled, it answered in 0.5 s. The edited file's row was marked, and the other file's row was not.

Deep scan deferred: no held-out corpus measurement or regression gate was run for this change.