Attribute --architecture edges by file when symbol names collide

[inth3shadows]

Summary

--architecture probes codegraph callees <name> for every enumerated symbol. That takes a bare name with no file disambiguation — and, measured against the pinned codegraph 1.5.0, it answers with the union of every same-named symbol's callees. So a name collision did not merely attribute an edge to the wrong file; it invented edges that exist in neither. On a purpose-built two-handle fixture, the pre-PR code drew 4 edges where only 2 were real.

TECHNICAL.md and USAGE.md both asserted codeshot had no way to fix this. That was false: codegraph node -f <file> <name> disambiguates a symbol to a file, and its trail line is file-qualified.

• probeFileEdges re-probes names appearing in more than one file via node -f; everything else keeps the cheaper callees --json route. node -f returns the symbol's full source on every call, so scoping it to the duplicate subset (~2% of symbols on a real ~1,900-node Go index) is what makes it affordable.
• Counting distinct files rather than symbol occurrences is load-bearing. Two same-named symbols inside one file have no file ambiguity at all — the bare-name union is already exactly right for them — and codegraph answers that case with two concatenated trail blocks, so routing it through the file-qualified probe would silently delete one symbol's edges.
• The trail is a human-facing summary, not an API. parseNodeCalls falls back to the bare-name probe — which over-reports rather than under-reports, the right way round to fail — on four measured conditions: no trail section (file-mode output, or a changed format); more than one trail section; a trail codegraph truncated with +N more (it caps at 12 entries — main in this repo has 23 callees and shows 12); or a reported **Location:** outside the requested file (-f is a preference, not a filter: node -f no/such/file.js -- buildDot still answers with render/callgraph.js's buildDot, exit 0).
• parseCodegraphOutput matches codegraph's not-found message against only the first non-empty line. Under the new json: false path the response embeds indexed source, and this repo's own test file contains that sentence verbatim.

Scope is deliberately partial. A colliding name with >12 callees falls back and keeps the old over-inclusive behavior. USAGE.md says so rather than implying it away.

Still unresolvable: duplicate file basenames (two index.js). unwrapQueryNodes keeps file nodes in the probed set, and node -f answers a file node in file mode. duplicateNameWarning now reports the resolved and unresolved halves separately.

Also folds in a one-line .gitignore fix: .codegraph//.ai/ matched directories only, so the shared-index symlink a worktree layout creates showed as permanently untracked.

Test plan

• node test/run.js → 108 passed, 0 failed (was 93 on master).
• New CLI regression test builds a fixture repo covering both collision kinds and asserts the real edges are present and the fabricated ones absent: a cross-file collision (two handles, must split) and a same-file one (two runs, must not). This repo's own index has zero duplicate names, so the existing self-test could never exercise either path.
• New unit tests pin each fallback condition: +N more → null, wrong file → null, file-mode output → null, multi-block → null, plus same-file duplicates keeping both edges and source containing the not-found sentence not being misread.
• Both committed diagrams re-verified against the exact CI guard (--check); unchanged, as expected for a repo with no duplicate names.