fix(trimmer): keep definition names and instance data in trimmed schemas

[LHMQ878]

Summary

Follow-up to #4036, which fixed ToolOutputTrimmer deleting tool parameters whose names collide with JSON Schema prose keywords and explicitly left the $defs case out of scope. This covers that case and one more of the same kind.

_trim_json_schema drops description, title, $comment and examples at every recursion depth. #4036 recognised that properties is keyed by user-chosen parameter names, not schema keywords, and special-cased it. Two groups are still walked as if they were subschemas:

1. Other name-keyed maps — $defs, definitions, patternProperties, dependentSchemas, dependentRequired. A definition named description is deleted while the $ref pointing at it survives, so the schema stops resolving. This is the case #4036 flagged as out of scope, and it is reachable through the SDK's own function_schema path, not just hand-written schemas — Pydantic keys $defs by class name, and a nested model class named description or examples produces exactly this:

$defs keys emitted by function_schema: ['description', 'examples']
after trim:                            []
dangling refs: [('note', '#/$defs/description'), ('samples', '#/$defs/examples')]

Through the public filter API (CallModelData, no network or API key), the trimmed schema is not merely lossy — it fails validation outright:

$defs in : ['Priority', 'description']
$defs out: ['Priority']
  #/$defs/description -> resolvable: False
jsonschema.Draft202012Validator(params).validate(...)
  -> PointerToNowhere: '/$defs/description' does not exist

patternProperties is affected whenever the pattern is an unanchored bare word ({"title": …} → {}); an anchored ^title$ happens to survive, which is why this is easy to miss.

2. Keywords holding instance data rather than subschemas — default, const, enum. Nothing inside these is a schema keyword, so a title key in a default value is part of the value:

default in : {'title': 'Untitled', 'description': 'auto', 'retries': 3}
default out: {'retries': 3}

The model is then shown a different default than the tool actually has. enum items and const objects lose the same keys.

The three categories are lifted into named module-level frozensets (_PROSE_SCHEMA_KEYWORDS, _NAME_KEYED_SCHEMA_MAPS, _DATA_SCHEMA_KEYWORDS) so the distinction is stated rather than implied by control flow, and adding a keyword later is a one-line change in one place. This only ever narrows what gets deleted — genuine subschema keywords such as items and propertyNames are still trimmed exactly as before, and no public API, signature or field order changes.

Test plan

Two cases added to tests/extensions/test_tool_output_trimmer.py, both driven through the public trimmer call rather than the private method:

• test_keeps_definition_names_and_instance_data_in_schema — every $ref still resolves, all five name-keyed maps keep their keys, default/const/enum are preserved byte for byte, and schema-level and nested prose is still stripped (so the trimming feature stays exercised).
• test_trims_prose_inside_genuine_subschema_keywords — a guard in the opposite direction: items and propertyNames must keep losing their prose, so the fix can't silently grow into a no-op.

Verified:

• tests/extensions/test_tool_output_trimmer.py: 41 passed with the fix; 1 failed, 40 passed with the src change reverted and the tests kept (fails on the $defs assertion, ['Priority'] != ['Priority', 'description']).
• tests/extensions/ (whole directory): 1105 passed, 40 skipped, 11 failed. I diffed the failing-test names against clean main (fc084ae2) — the two sets are identical, so there is no regression. They are pre-existing sandbox/session failures (test_bootstrap_persistent_resources_*, test_query_run_loop_*, test_get_items_*, test_runner_with_session_settings_override, test_run_loop_extension_reexports_cloud_bucket_strategy).
• ruff check: all checks passed. ruff format --check: 2 files already formatted.
• mypy on the two changed files: the only 2 errors are in sandbox/util/tar_utils.py:161 and extensions/sandbox/modal/sandbox.py:126; I confirmed both reproduce with my changes stashed, so they are pre-existing and untouched by this PR.

Verification-script note: .agents/skills/code-change-verification/scripts/run.sh could not run here — make is unavailable on this machine — so I ran the underlying targets directly.

Issue number

None — self-reported. This is the follow-up #4036 described as "a one-line follow-up ... if you'd rather have it covered here"; the default/const/enum half was not part of that description.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh — could not: make unavailable locally (see Test plan); ran the equivalent targets directly instead
• I've confirmed all verification steps pass (format, lint, typecheck and the affected tests; remaining failures verified pre-existing on main by diffing failure sets)
• If using Codex, I've run /review before submitting this PR — n/a

AI assistance was used to draft this change; it was reviewed, reproduced and verified by me.