fix: restore incremental Responses reasoning

[sandexzx]

Summary

Restore incremental reasoning on Command Code's Provider API Responses wire, reproduced with deepseek/deepseek-v4.1-flash-fast.

After #133 selected Responses for models advertising /responses, Command Code's response.reasoning.delta events stopped reaching Pi as thinking deltas. pi-ai 1.0.3 understands response.reasoning_text.delta, but not this non-standard alias. The final response.output_item.done still supplies the complete reasoning text, which explains why thinking appeared all at once at the end.

Changes

• Normalize only valid response.reasoning.delta SSE payloads to response.reasoning_text.delta, including the matching SSE event: name, before delegating to pi-ai.
• Apply the wrapper only after resolving the model's Responses wire; leave Chat Completions, Anthropic Messages, generate fallback, and 403 upgrade_required detection unchanged.
• Process complete SSE frames incrementally, preserving UTF-8 boundaries, a leading BOM, LF/CRLF/CR framing, multiline data, cancellation, and input stream errors.
• Leave standard events, malformed payloads, completion events, and persisted reasoning signatures unchanged. No model allowlist, dependency changes, or pi-ai patch is required.
• Add normalizer tests and a real-Pi regression whose mock server cannot finish the reasoning item until both thinking deltas have reached the client. Assert the final text is not duplicated.
• Await RPC child-process closure before deleting test session directories, fixing the observed intermittent ENOTEMPTY cleanup failure. Report PASS only after cleanup succeeds.

Validation

• npm test, with PI_LOCAL_REQUIRED=1 and OMP_COMPAT_REQUIRED=1 and a temporary OMP installation supplied through OMP_BIN: all suites passed, including both real hosts against mock APIs.
• npm run format:check.
• git diff --check.
• Three additional consecutive npm run test:pi-local runs: all exited successfully after cleanup.
• npm pack --ignore-scripts followed by npm run test:release-package -- <tarball>, with OMP_BIN: the installed tarball passed the Pi and OMP mock suites.
• I verified the fix in Pi with a live Command Code request to DeepSeek V4.1 Flash Fast: reasoning now streams incrementally instead of appearing all at once at the end.

[pierreraby]

Verified: real bug, confirmed live — the gateway emits response.reasoning.delta (37x on deepseek-v4.1-flash-fast, 0x canonical alias) and pi-ai 1.0.3 drops them, so thinking lands all-at-once on every Responses-routed model. I captured raw frames: {type, sequence_number, item_id, output_index: 0, content_index, delta: string} — integer output_index + string delta, so the normalizer's schema assumption holds against the real gateway, not just the mock.

Local on 7e0e831: typecheck, test-transport 13/13, test-models 28/28, test-pricing 8/8, pi-local PASS, prettier clean. A fresh-context reviewer also checked the SSE parsing, wrapper scoping (Responses-only, 403 fallback preserved), and streaming properties — sound.

Two non-blocking P2s for follow-up: (1) finish() emits a trailing unterminated frame raw instead of normalizing it; (2) on schema mismatch the frame passes through silently, reverting to all-at-once with no signal. Also: this PR's wait-for-close cleanup and main's retry loop (0.7.5) fix the same ENOTEMPTY flake from opposite ends — suggest keeping wait-for-close as primary plus a longer retry backstop (and extending the waiter to runPi's timeout path) rather than picking one.

Leaving the merge to the maintainers.

[pierreraby]

Re-verified on 8a615be (merge of main + P2 hardening): typecheck, test-transport 16/16, pi-local PASS, prettier clean. The finish() normalization and schema-mismatch warning address my earlier P2s; wait-for-close + retry loop are combined as suggested. Merging once CI is green.