feat(types): regenerate schemas + inject Literal-discriminator defaults

[bokelley]

Summary

Upstream spec added a cluster of breaking changes — most Asset classes went from "discriminator is implicit" to "required tag field with single-valued Literal[...] type". Pydantic honors required literally, breaking ergonomic construction:

# Before upstream change:
TextContent(content="hello")

# After upstream change, pre-fix:
TextContent(asset_type="text", content="hello")  # must repeat the tag

Our v3→v4 migration already handles the <Type>Asset → <Type>Content rename; this new breaker is separate (required tag on the renamed class) and hits 15 test failures across test_client, test_preview_html, and test_discriminated_unions.

The fix: pattern-based default injection

inject_literal_discriminator_defaults() in scripts/post_generate_fixes.py. Walks the AST of each generated Pydantic model; for any AnnAssign whose annotation is Literal['x'] (bare or Annotated-wrapped) and whose value is None (no default), appends = 'x' on the same line via a line-based edit.

Touched 687 fields across 615 classes in 105 files — not just Asset types. Pricing options, webhook types, catalog types, and every other spec discriminator with a single-value Literal all get defaults.

Why this is safe

• Wire consumption is unchanged. Dicts with the tag present validate exactly as before.
• Validation strength is preserved. The Literal type still rejects any other explicit value — defaulting doesn't weaken the tag check.
• Pattern-based, not value-specific. Robust to spec-value churn (e.g., if upstream renames asset_type values) as long as the single-value-Literal shape holds.

What's NOT fixed

The 3 WebhookAsset / CatalogAsset / VAST residual failures I flagged in the assessment turned out to also be Literal discriminators — all 15 failures are resolved by this one change. Future genuine-non-discriminator breakers (where a field became required with no canonical default) would need a separate mitigation (BeforeValidator coercion, better error messages) — this PR doesn't add that infrastructure because we don't need it yet.

Tested

• +16 tests in tests/test_literal_discriminator_defaults.py:
  • Asset-content defaults (5 types × asset_type check)
  • VAST/DAAST dual-discriminator defaults (3 shapes)
  • Validation-strength preservation (wrong explicit tag still raises ValidationError)
  • Wire-format compatibility (dict → model with and without tag)
  • Injector meta-tests (skips multi-value Literals, extracts single-value, skips non-Literal)
• Full suite: 2102 passed, 22 skipped (was 2086 on previous green; 15 new cleared + 16 test additions - 15 removals).
• ruff, mypy clean across 678 source files.

Also in this PR (upstream pulls)

• docs(design): FastMCP native registration investigation (#209) #250 (design doc): FastMCP native-registration investigation
• feat(a2a): auto-retain contextId + taskId across multi-turn calls #251: A2A contextId + taskId auto-retain across multi-turn calls (+22 integration tests)

Test plan

• CI green across Python 3.10–3.13
• Review scripts/post_generate_fixes.py:inject_literal_discriminator_defaults — the AST walk + line edit is the risky part. Edit order is sorted descending so earlier edits don't shift later line indices.
• Review a sampling of the regenerated Pydantic classes — confirm the = 'text' suffix is well-formatted and syntactically valid (easy to spot-check on src/adcp/types/generated_poc/core/assets/text_asset.py).
• Optional: consider whether to surface the injection count in the regen summary output (currently just logged as a count line — quiet, can go unnoticed).

🤖 Generated with Claude Code