fix(schema): harden _schema_short_hash to strict JSON serialization

[yeongseon]

Context

Split from #379 (was "item B"). This is defensive hardening, not a live regression — schemas always arrive as pure JSON from Pydantic model_json_schema(), so the anonymous-hash path is not reached with non-JSON input in practice. Tracked separately so it does not dilute the JSON-Pointer-safety fix in #379.

_schema_short_hash() (utils.py:144) uses default=str:

canonical = json.dumps(schema, sort_keys=True, default=str)

default=str silently stringifies non-JSON objects. str(obj) may embed a memory address, making InlineSchema_<hash> differ across processes — contradicting the "stable/deterministic hash" docstring — and it masks a violation of the endpoint-metadata contract (schemas must be plain JSON-compatible dicts).

Acceptance Checklist

• Replace default=str with strict serialization: json.dumps(schema, sort_keys=True, separators=(",", ":")) so non-JSON input raises TypeError loudly instead of producing a non-deterministic name.
• Test: a non-JSON-serializable schema raises TypeError; identical schemas hash identically.
• Coverage stays >= 95%.

Out of scope

• JSON-Pointer-safe component names (fix(schema): make hoisted component names JSON-Pointer safe and hash strictly #379).

References

• src/azure_functions_openapi/utils.py:142-145
• fix(schema): make hoisted component names JSON-Pointer safe and hash strictly #379 (parent), PR fix(schema): make hoisted component names JSON-Pointer safe and hash strictly #383

[yeongseon]

Resolved by #383 (strict _schema_short_hash, split out from original issue B). Closing manually — GitHub's gh pr edit GraphQL failure on this repo dropped the auto-close link from the PR body.