perf: extend _returns_model caching to pos_handler, async_pos_handler, fast_model_handler

[justrach]

Closes #157. Sub-task of #42. Follow-up to #156 (now merged into `release/1.0.30`). Re-opens replacing the original PR #160 which auto-closed when its stacked base branch was deleted on #156's merge.

Summary

PR #156 introduced _returns_model(handler) and consulted it at dispatch in three fast-handler sites. The same per-response hasattr(result, "model_dump") pattern existed in three more sites that #156's scope deliberately left out. This PR closes the loop.

line	site	function
~1268	sync positional dispatch	create_pos_handler.pos_handler
~1297	async positional dispatch	create_async_pos_handler.pos_handler
~1559	model-input dispatch	create_fast_model_handler.fast_model_handler

Each gets:

1. _returns_md = _returns_model(original_handler) bound once at the top of the factory.
2. Per-response check switches from if hasattr(result, "model_dump"): ... to if _returns_md or (_returns_md is None and hasattr(result, "model_dump")): ....

create_fast_model_handler is the most interesting site: it already takes a model_class for input validation, but the output shape is still unknown until you check the user handler's return annotation. The same helper applies — we introspect original_handler (the user's handler), not model_class.

Identity-preserving for all three branches (True / False / None). The _returns_md is None fallback keeps the per-response hasattr behavior for unannotated handlers.

Files / subsystems touched

• python/turboapi/request_handler.py — 3 closure-time bindings + 3 dispatch-site swaps (one per factory function).
• tests/test_returns_model_cache_extended.py — new file. 10 targeted tests, 3-4 per handler, covering each annotation shape (-> Model, -> dict, unannotated).
• benchmarks/bench_model_dump_cache_extended.py — new isolated microbench targeting create_pos_handler (the cheapest of the three; the other two share the same dispatch code paths and benefit identically).

No public API change. No dependency change. No uv.lock change.

Red-to-green

Microbench (load-bearing artifact)

.venv/bin/python benchmarks/bench_model_dump_cache_extended.py, n=200k × 5 runs, free-threaded 3.14t, M-series macOS:

```
case pre median post median delta pre p99 post p99 delta
leaf_dict 668.6 ns 654.3 ns -2.1% 875 ns 875 ns 0%
model_dhi 2711.2 ns 2645.5 ns -2.4% 3125 ns 2917 ns -7%
unannot 686.8 ns 669.4 ns -2.5% 959 ns 958 ns noise
```

Modest in absolute terms (the per-call savings is one hasattr ≈ 30 ns) but consistent across all three shapes. The model-return path's p99 drops 7%, matching the pattern from #156's bench.

Pre-PR baseline collected by `git stash`-ing the patch file (with #156's _returns_model helper still in place) and re-running the same bench against the unpatched dispatch sites.

Targeted tests

```
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/python -m pytest -q \
tests/test_returns_model_cache.py \
tests/test_returns_model_cache_extended.py \
tests/test_async_handlers.py \
tests/test_post_body_parsing.py
```

Result: 32 passed in 24.48s (10 new for #157 + 10 from #156's test file + 12 existing). Pre-existing `PytestReturnNotNoneWarning`s unchanged by this PR.

Risk

Low. Same shape as #156. The new test file directly exercises:

• `create_pos_handler` × 4 shapes (model annotated, dict annotated, unannotated returning model, unannotated returning dict)
• `create_async_pos_handler` × 3 shapes
• `create_fast_model_handler` × 3 shapes (model in / model out, model in / dict out, model in / unannotated returning model)

All confirm: annotated-as-model still calls model_dump, annotated-as-dict skips it correctly, unannotated still falls back to hasattr.

🤖 Generated with Claude Code

---

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 07d8df9ce8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve tuple responses before cached model dumping

When a handler is annotated as returning a model but returns a status tuple such as ({'detail': 'created'}, 201) or an error tuple on some branch, _returns_md is True, so this now calls model_dump() on the tuple and the wrapper returns a 500 instead of reaching the existing tuple-response branch. The same ordering was added in create_async_pos_handler and create_fast_model_handler; tuple handling needs to remain before the cached model-dump shortcut, as in the no-args fast path.

Useful? React with 👍 / 👎.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.