fix(server): centralize pre-validation hook composition

[sangilish]

Summary

Refs #859.

This is a focused follow-up to the hook-composition work that already landed in #860. Current main already supports ordered pre-validation hook chains, so this PR avoids reimplementing that behavior and only tightens the remaining pieces from the #859 brief and #860 review notes:

• Moves shared pre-validation hook types, composition, flattening, and application into adcp.server._hooks.
• Keeps adcp.server.spec_compat backward-compatible by re-exporting the public hook types/helper.
• Makes hook-chain failures surface the failing chain index and callable name in the INVALID_REQUEST message.
• Replaces mock-only MCP forwarding coverage with a real MCPToolSet.call_tool(...) behavior test.
• Adds real A2A executor coverage for ordered hook chains.

Review note

I opened this as a draft because #860 already shipped most of #859. The main question is whether this small cleanup/diagnostic follow-up is still useful, or whether #859 should be closed as already covered by #860.

Current CI note

The draft currently has red storyboard checks against the moving adcp-3.1 npm dist-tag. The failing PR run installed @adcp/sdk@8.1.0-beta.13; the latest passing main CI run I checked used @adcp/sdk@8.1.0-beta.12. I also reproduced the same classes of storyboard failure from a detached current-main worktree with ADCP_SDK_VERSION=adcp-3.1, including the reference seller 3.1 storyboard and proposal-finalize recovery expectation.

I am therefore keeping this as draft rather than marking it ready while that runner/schema drift is active.

Testing

• uv run pytest tests/test_pre_validation_hooks.py tests/test_a2a_server.py -q
• uv run pytest tests/test_pre_validation_hooks.py tests/test_a2a_server.py tests/test_spec_compat_hooks.py tests/test_spec_compat_hooks_deprecation.py tests/test_unknown_field_policy_server.py tests/test_schema_validation_server.py tests/test_serve_validation_passthrough.py tests/test_testing_decisioning.py tests/test_public_api.py -q
• uv run ruff check src/adcp/server/_hooks.py src/adcp/server/spec_compat.py src/adcp/server/mcp_tools.py src/adcp/server/a2a_server.py src/adcp/server/serve.py src/adcp/server/__init__.py tests/test_pre_validation_hooks.py tests/test_a2a_server.py
• uv run mypy src/adcp/server
• uv run mypy src/adcp
• uv run pytest -q
• git diff --check

Full uv run ruff check still fails on unrelated existing test lint drift outside this PR's touched files, including long lines and unused imports in existing tests.

[bokelley]

Decision: this is worth landing.

#860 already shipped the core of #859 — PreValidationHook / PreValidationHookChain / compose_pre_validation_hooks are live and threaded through MCP + A2A + serve(). So this PR is the diagnostics + test-quality follow-up, and that part earns its place:

• named-chain error attribution (the failing chain index + callable surfaced in INVALID_REQUEST messages, instead of the generic "pre_validation_hook raised ..."),
• the swap from mock/spy MCPToolSet tests to a real call_tool() behavior test, plus the A2A ordered-chain executor test and chain-failure-attribution tests,
• consolidating the hook types/composition into adcp/server/_hooks.py as a single source of truth (spec_compat.py re-exports).

To land:

1. Rebase onto current main (the PR is CONFLICTING/DIRTY).
2. Re-run CI on the current @adcp/sdk adcp-3.1 dist-tag — the red storyboard checks looked like dist-tag drift (the failing run used @adcp/sdk@8.1.0-beta.13), reproducible on a clean main worktree; confirm they're unrelated to this diff.
3. Since the symbols move from spec_compat.py to _hooks.py, confirm tests/test_import_layering.py and test_public_api.py stay green.

@sangilish — can you rebase? Happy to take it over and push the rebase if you'd prefer. Keeping #859 open as the tracker until this lands.

[bokelley]

Rebased onto current main and pushed to claude/issue-859-pre-validation-hooks-rebase for review/fast-forward.

Conflict resolution summary:

• src/adcp/server/serve.py: The PR's base had StreamableHTTPSessionManager imported at the top-level; current main replaced that with ADCPStreamableHTTPSessionManager from adcp.server.mcp_sessions. Resolution: dropped the stale StreamableHTTPSessionManager line, kept the from adcp.server._hooks import PreValidationHooks addition.
• tests/test_a2a_server.py: Both main (test_execute_skips_scalar_datapart_and_uses_text_fallback) and this PR (test_pre_validation_hook_chain_runs_through_a2a_executor) inserted tests at the same point after test_execute_with_datapart. Resolution: both tests included in order.

Checks on the rebased branch (Python 3.11):

• pytest tests/test_pre_validation_hooks.py tests/test_a2a_server.py tests/test_spec_compat_hooks.py tests/test_spec_compat_hooks_deprecation.py tests/test_import_layering.py tests/test_public_api.py → 128 passed
• ruff check on touched files → all checks passed
• mypy src/adcp/server → no issues found in 31 source files

The storyboard CI failures match the dist-tag drift @sangilish described and are reproducible from a clean main worktree — unrelated to this diff.

@bokelley — the rebased branch is ready. You can fast-forward feat/pre-validation-hooks-859 from it, or redirect the PR to merge from claude/issue-859-pre-validation-hooks-rebase directly.

---

Generated by Claude Code

[bokelley]

Superseded by #938, which re-applies this PR's diagnostics — the named-chain error attribution (pre_validation_hook[i] ...) and the real-path MCP/A2A tests, consolidated into a new adcp/server/_hooks.py — freshly on current main (#930/#933 had moved the server files this branch was based on, so a rebase wasn't clean). You're credited as co-author on #938, @sangilish. The core hook composition shipped earlier in #860, and #938 closes #859. Thanks for the original work.