fix(parse): classify every spawn candidate as InvokeSpawn (calque#189 follow-up)

[scttfrdmn]

Summary

This is a follow-up to #190 (calque#189), prompted by a direct question: "why not synthesize a test that exercises what the AI-Almanac script is doing under controlled and understood conditions?" I hadn't — I'd only tested SpawnCallSites in isolation. Writing the actual end-to-end pipeline test surfaced a real second bug.

• fix(parse): recognize .spawn() on a dict-of-functions Subscript (calque#189) #190's fix made a dict-subscript .spawn() call site visible (SpawnCallSites now emits a SpawnCallSite{Target: "bundle", ...} instead of nothing).
• But invocationKinds' consider() call for the "spawn" case was still keyed on the empty ic.Target, not the candidates — so the candidate callable's own ir.Function.Invoke never became InvokeSpawn. ResolveSpawnCallables (internal/exec/spawnshard.go) only ever finds a callable whose OWN Invoke is InvokeSpawn, so it returned zero callables, and BuildSpawnManifests' byTarget lookup had a call site with nothing to shard against — zero shards, silently. The exact failure .spawn() call site via a dynamic subscript (dict[key].spawn(...)) resolves to an empty target — spawnrun can't drive it #189 exists to prevent, one layer deeper.
• invocationKinds now classifies every candidate as InvokeSpawn when a spawn call site's Target is empty but Candidates is populated.

Why the original PR missed this

The original fix touched two layers (SpawnCallSites for call-site resolution) but I only unit-tested that layer in isolation, plus one subprocess-level test confirming the JSON wire shape. I never assembled the real pipeline (Parse → ResolveSpawnCallables → SpawnCallSitesReport → BuildSpawnManifests) end-to-end — which is exactly what spawnRunFromScript (cmd/calque/spawnrun.go) actually does. The defect lived in the gap between two already-individually-tested layers, invisible to either one's own unit tests.

What this PR adds

• The classification fix itself (internal/parse/parse.go).
• TestInvocationKindsClassifiesSpawnCandidatesNotEmptyTarget — pure unit test on invocationKinds.
• cmd/calque/spawn_dict_dispatch_e2e_test.go's TestSpawnDictDispatchEndToEndProducesRealShards — a real end-to-end test wiring the full pipeline together against the spawn_dict_dispatch.py fixture, asserting a real, non-empty shard comes out the other end. Verified this test actually fails without the fix (reverted parse.go locally, confirmed both assertions fail with the exact "zero callables" symptom, then restored the fix).

Also found, filed separately (not fixed here)

calque#191: spawn-run only ever binds a spawned callable's FIRST positional arg regardless of its real arity (found via the same end-to-end test — bundle's job_id/config args showed only job_id bound in the built shard's payload). This is a distinct, pre-existing limitation in spawnshard.go's own binding protocol, out of scope for #189's target-resolution fix — calque#187's arity guard doesn't cover spawn-run's separate driver at all.

Test plan

• go build ./... && go vet ./... && gofmt -l . && go test -count=1 ./... — all pass
• golangci-lint run ./... — 0 issues
• ruff check . — all checks passed
• Confirmed TestSpawnDictDispatchEndToEndProducesRealShards fails without this fix (manually reverted parse.go, re-ran, confirmed the exact "0 callables / 0 shards" failure), passes with it