fix(idempotency): support @wrap on arg-projected methods (update_media_buy)

[bokelley]

Closes #559

Summary

@IdempotencyStore.wrap raised TypeError: missing required positional argument: 'params' when applied to methods the framework dispatches via the arg-projector path (e.g. update_media_buy). The arg-projector calls method(media_buy_id=..., patch=..., ctx=ctx) with no positional params — so the wrapper's params: Any parameter was never filled.

Root cause: the wrapper had no way to obtain the canonical wire-shape Pydantic model needed for RFC 8785 hashing and idempotency_key extraction, since the projected kwargs are a field-subset and cannot be reconstituted into a valid model.

Fix approach: dispatch now threads the original params object as __adcp_params__=params when is_wrapped(method) is True, giving the wrapper the full model without touching the normal call site. The wrapper pops the sentinel, detects the arg-projector convention (params is None and len(field_kwargs) > 0), and uses the canonical model for hashing while forwarding the original projected kwargs to the handler.

Edge case fixed by pre-PR review: direct adopter unit-test calls (e.g. seller.update_media_buy(media_buy_id=..., ctx=ctx) without dispatch or sentinel) previously would have fallen into _prepare(None, ...) and raised a confusing TypeError inside _to_dict. Now the wrapper falls through to the handler when arg_projected=True and the sentinel is absent, so wrapped methods are unit-testable without going through dispatch.

Changes

• src/adcp/server/idempotency/store.py — _MISSING sentinel, updated _wrapped signature (params: Any = None), arg-projector detection, sentinel pop, sentinel-absent fall-through, kwargs.clear() for clarity
• src/adcp/decisioning/dispatch.py — is_wrapped as _is_idem_wrapped import; conditional __adcp_params__=params injection on both async and sync executor paths; inline comment on the sync path (forward-compat only, since wrap always returns an async def)
• tests/test_server_idempotency.py — 5 new tests in TestArgProjectedCallingConvention: no-TypeError, cache-hit dedup, payload conflict, no-key fall-through, and direct-call-without-sentinel fall-through

Tested

• pytest tests/test_server_idempotency.py — 54 passed (5 new regression tests)
• pytest tests/ (full suite) — 3,940 passed, 1 pre-existing failure (TLS network test unrelated to this change)
• mypy src/adcp/ — 0 errors
• ruff check src/ — 0 errors

Pre-PR review

• code-reviewer: approved after blocker fix — blocker (direct-call without sentinel → confusing TypeError in _prepare) resolved by sentinel-absent fall-through; kwargs = {} → kwargs.clear() for clarity; sync executor dead path documented with inline comment. 1 nit surfaced below.
• ad-tech-protocol-expert: approved — non-breaking per AdCP #2315; RFC 8785 hash equivalence preserved (full wire-shape model hashed, not projected subset); per-principal scope isolation and context-echo contract unaffected; idempotency_key extraction from canonical model is sound.

Nit (not blocking): __adcp_params__ is not a reserved kwarg at the Pydantic level — a request model with a field named __adcp_params__ would have its field value shadowed at the call site by the sentinel. The double-underscore prefix is a reasonable deterrent; the sentinel name is documented in store.py's module docstring as reserved in the arg-projector calling convention.

---

Triage-managed PR. This bot does not currently iterate on
review comments or PR conversation threads (only on the source
issue). To unblock:

• Push fixup commits directly: gh pr checkout 560 →
  fix → push.
• Or re-trigger: comment /triage execute on the source
  issue.

See adcp#3121
for context.

Session: https://claude.ai/code/session_01HopGAcSiNk5SWbLpE9fd6b

---

Generated by Claude Code

[bokelley]

Superseded by #567 (merged today as 1abdae6), which closed #559 with the wrap-side resolution architecture rather than the dispatch-side sentinel approach this draft proposed. Per CLAUDE.md's parallel-race guidance for stale claude-triaged draft PRs: closing rather than rebasing, since the two architectures aren't compatible (#567 explicitly chose against threading dispatch→idempotency coupling). Issue #559 already in 4.4.0 changelog via #567.

[bokelley]

Noted — supersedence by #567 (1abdae6) confirmed. Issue #559 is closed via the wrap-side resolution. No further action on this draft.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01Rvm9f1JFJLjBfyvVrXVe1A

---

Generated by Claude Code