fix(run): refuse a multi-arg non-starmap warm unit instead of silently NameError'ing (calque#187)

[scttfrdmn]

Summary

• Found while re-verifying calque#79's closing claim that AI-Almanac's three real scripts run end-to-end against calque v0.6.0: a picked warm unit with 2+ non-self/cls positional args that isn't .starmap()'d silently NameError'd on every synthetic item, since dryRunWarm/manifestBodyForUnit only ever bind the FIRST positional arg outside the .starmap() splat path.
• Real-world hit: forecasts_app.py's run_forecast_inference(job_id, model_id, config) is invoked via .spawn(), not .map()/.starmap() — checkInvokeSupport had no arity check for anything besides InvokeStarmap/InvokeForEach, so this fell through silently.
• checkInvokeSupport now refuses loudly with a clear message naming the function, its arg count/names, and pointing at calque#187, instead of a confusing downstream NameError.
• Critical carve-out: calque real --arg-file/--arg-json already solves this correctly for a real run when the caller supplies a real per-position tuple (e.g. app.py's run_benchmark_local, calque#178's real-hardware-verified path, invoked via .remote() with 4 args) — a blanket refusal would have broken that already-working, already-verified case. Added a hasRealArgTuple param to checkInvokeSupport, threaded through only at realrun.go's own call site (forceStarmap), so that path is unaffected.

Files touched

• cmd/calque/run.go — checkInvokeSupport's new arity guard + hasRealArgTuple param.
• cmd/calque/realrun.go — passes forceStarmap (true when --arg-file/--arg-json supplied a real tuple).
• cmd/calque/fleetrun.go — passes false (fleet has no --arg-file equivalent).
• cmd/calque/gate_test.go — updated existing call sites, new tests (TestCheckInvokeSupportMultiArgNonStarmapRefuses, TestCheckInvokeSupportMultiArgWithRealArgTupleFine, TestCheckInvokeSupportSingleArgNonStarmapFine).
• cmd/calque/starmap_e2e_test.go — updated call site.
• CHANGELOG.md.

Test plan

• go build ./... && go vet ./... && gofmt -l . && go test -count=1 ./... — all pass
• golangci-lint run ./... — 0 issues
• ruff check . — all checks passed
• Manually re-verified against fresh clones of all three real AI-Almanac scripts:
  • forecasts_app.py / blending_app.py now refuse honestly and immediately (previously crashed 50x on an unrelated missing-import/missing-secret error before ever reaching the unbound-arg bug)
  • app.py's dry-run path (no --arg-file) now also correctly refuses run_benchmark's 3-arg signature (previously masked by an earlier SERVICE_ACCOUNT_JSON KeyError crash)
  • Confirmed the carve-out doesn't regress calque#178's real-hardware-verified run_benchmark_local path (new regression test TestCheckInvokeSupportMultiArgWithRealArgTupleFine)