fix(embed): reject impossible entry counts before allocation

[redxzeta]

Corrupt embedding headers currently reserve memory before validating that entries fit in the file. The existing combined-documents fixture corrupt requests 92,164,907,664 bytes and aborts on clean upstream cf82f5d (Fixes #80).

Validate the count against the minimum encoded entry size before reserving. The format and valid empty-ID/empty-vector cases remain unchanged. New regressions cover corrupt, a maximum count, a truncated entry, and valid round trips.

Validation: formatting and diff checks pass; both new embedding-format tests and all five existing combined-documents tests pass. This is a separate prerequisite for diagnostics PR #81. No checks are disabled. No deployment or activation.

Hosted CI run 37177784672 for head 2c621c2 reports action_required. The run page explicitly says it awaits maintainer approval; the API reports zero jobs, with no hosted validation result. The authenticated contributor account has read-only upstream permissions and cannot approve the run. Required PR review and successful hosted CI remain pending.

[redxzeta]

Investigated CI run 37177784672 on head 2c621c25d470cd37cb54a15d0775c03d0a08294d. GitHub explicitly reports that the workflow awaits maintainer approval. No jobs started and no logs are available, so this is an approval gate with no hosted test/build result.

An upstream maintainer must approve the run; the authenticated contributor account has read-only upstream permissions. No source or workflow change, check suppression, or rerun was performed. The focused local validation remains as recorded above; required review and successful final-head hosted CI remain pending. T3 watching remains enabled.