fix(codegen): rewrite compact ConfigDict extra in place

[bokelley]

Summary

Fixes the one remaining gap in #907: _set_class_extra_allow in
scripts/post_generate_fixes.py only matched the multi-line
model_config = ConfigDict(\n ...) form. A compact single-line
model_config = ConfigDict(extra='forbid') did not match, so it fell
through to the else branch that prepends a fresh model_config block —
producing two model_config declarations in the same class, which breaks
Pydantic at import time.

The other #907 sub-items (missing ConfigDict import injection, enum-base
exclusion for the title-less root sentinel, the #902 short-circuit note,
docs follow-ups) are already shipped. This PR closes the compact-ConfigDict
residual only.

Repro (before fix)

A compact ConfigDict(extra='forbid') on a class that should get
extra='allow' produced a duplicate:

class PartnerPayload(AdCPBaseModel):
    model_config = ConfigDict(
        extra='allow',
    )
    model_config = ConfigDict(extra='forbid')   # duplicate — last one wins
    name: str

Change

_set_class_extra_allow now matches the compact single-line form
(ConfigDict(<args>) where the open paren is not followed by a newline) and
rewrites it in place:

• extra='allow' already present → returns "already", no change
• extra='forbid'/'ignore' → flipped to extra='allow' in place
• other args, no extra= → extra='allow' inserted into the call
• empty args → extra='allow' inserted

Exactly one model_config is emitted; the multi-line and no-config paths are
unchanged.

Tests

Added test_compact_configdict_forbid_rewritten_in_place to
tests/test_extra_policy.py asserting the compact extra='forbid' →
extra='allow' rewrite and that no duplicate model_config is produced.

Verification

• uv run --extra dev pytest tests/test_extra_policy.py -q → 14 passed
• make lint → passed
• make typecheck → passed
• make validate-generated → passed
• No regen needed; no generated files changed.

🤖 Generated with Claude Code

[aao-ipr-bot]

Clean fix. The compact and multi-line forms are now mutually exclusive arms over the same model_config, so exactly one declaration survives — which is the whole point.

Things I checked

• No multi-line mis-fire. compact_pattern group 2 is [^\n]*? — it can't cross a newline, so a true ConfigDict(\n ... \n) (open paren immediately followed by \n) never matches it. And compact_match is only consumed in the elif, reachable only when config_match is None. The multi-line path always wins when present. scripts/post_generate_fixes.py:516-517, :542.
• In-place rewrite, single model_config. Reconstruction splices body[:start] + group(1) + new_args + group(3) + body[end:] — rewrite in place, no prepended second block. :556-563. The new test's updated.count("model_config") == 1 locks it.
• Branch coverage. extra='allow' present → \"already\"; forbid/ignore → re.sub count=1; other args → \"extra='allow', \" + args; empty → extra='allow'. All four produce valid Python (ConfigDict(extra='allow', populate_by_name=True), ConfigDict(extra='allow')). :545-554.
• Quote handling. extra=(['\"])(?:forbid|ignore)\1 matches single or double quotes; replacement normalizes to single-quote extra='allow', same as the existing multi-line arm at :521-527. Consistent.
• Scope. body is bounded to one class by the (?=^class |\Z) lookahead, so .search only sees the target class's own config. No cross-class bleed.
• Layering. Build-time codegen post-processor — no src/adcp/ public surface, no wire types, no generated_poc/ hand-edits. code-reviewer: clean, zero findings (Blocker/Major/Minor all none). Commit prefix fix(codegen) is the right semver signal — non-breaking, no public export touched.

Minor nits (non-blocking)

1. Live test run not executed here. Sandbox blocked pytest/branch checkout, so the rewrite is verified by static analysis of the regex + reconstruction, not a live run. The author reports 14/14 plus lint/typecheck/validate-generated green; CI is the backstop. No concern — noting it for honesty.

LGTM.