test: judge the package nodes annot and crossref already rank first

[tae2089]

Closes #102. Rebased onto main after #101, #107 and #104 landed; the numbers below are remeasured on that base.

What this is

A re-judgment of the golden set, with no code beside it. Two queries name a package by the name that package declares and were judged on everything but the package node itself.

• annot — prefix query. Its key already accepts a prefix match on a name (graph.Annotation is judged though the query stops five letters short), and internal/domain/annotation declares package annotation.
• crossref — one word naming a whole feature. internal/app/crossref declares package crossref, and impact was judged on that same reasoning already.
• mcp gains no node. Its why now records why the other package node in its pool stays unjudged: internal/runtime/mcp declares package mcpruntime, and only a directory in its path is spelled mcp.

A sweep of all four corpora found no other unjudged package node in any top-15 pool: cobra and context-diary have none in any pool at all, and gorm's clause is already judged.

Movement

Measured by rerunning the report with each edit alone:

annot     relevant 4 -> 5, found 2 -> 3, first relevant rank 2 -> 1
crossref  relevant 3 -> 4, found 3 -> 4, rank 1 unchanged
mcp       unchanged

Only the prefix bucket moves: 0.882 (15/17) top1 4 MRR 0.900 becomes 0.895 (17/19) top1 5 MRR 1.000 — the same movement as before the rebase; that bucket is untouched by the three PRs that landed in between.

ccg ALL         0.744 (125/168) 47 62 0.644  ->  0.747 (127/170) 48 62 0.649
ccg ANSWERABLE  0.838 (124/148) 46 61 0.697  ->  0.840 (126/150) 47 61 0.703

cobra, context-diary and gorm are untouched; no fixture was recaptured. testdata/README.md's ## The two totals block is updated to the new output, so it does not go stale again.

The top1 annot regains is the one it lost in #101 when package nodes entered the pool. The annotation package node is what the ranker had already put at rank 1 while the key called it wrong. Nothing about the ranking changed here — only which nodes the key admits are right.

Rule

testdata/README.md gains the rule these three judgments follow, with its two negative cases: a path segment spelled like the query is not the package's name, and containing an already-judged file is corroboration after the name matched, never a reason on its own.

Verification

• CGO_ENABLED=1 go test -tags "fts5" ./... -count=1 — all packages pass on the rebased base
• go vet ./... — clean
• gofmt -l . — only the two pre-existing parse fixtures under testdata/binding_gap/go/
• baseline.json re-recorded with -update-golden; the diff is the four numbers above and nothing else

🤖 Generated with Claude Code