fix: surface token parse failures as request errors

[james-bg]

Split out of #1765 per the sequencing discussed there — this is the part accepted as a pure bugfix.

The token parser's internal stream (Readable.from(StreamParser.parseTokens(...))) has no 'error' listener, so a parse failure inside a response surfaces as an unhandled 'error' event and crashes the process instead of failing the request.

This PR re-emits stream errors from Parser, and routes them in SentClientRequest through the existing socket-error path: the active request fails with the parse error and the connection is closed — a parse failure leaves the connection at an undefined position in the TDS stream, so it cannot be recovered at the request level. The listener is removed at end-of-message; during a canceled response's drain it stays attached, so a parse failure there is still surfaced.

Complements #1759, which fixed the same unhandled-'error' hole for the login path.

Test: connection-parse-error-test.ts — a scripted server answers a SQL batch with an invalid token type (0x00). On current master the test dies with an uncaught Unknown type: 0 stream error; with this fix the request errors and the process stays alive. Full unit suite and lint pass.

The inactivity-timer half of #1765 will follow separately, targeted at v21 on top of the AbortSignal work, per the maintainer's sequencing.

[arthurschreiber]

Thanks for splitting this out. The core change is correct and the test covers it: with the src changes reverted the new test dies with an uncaught Unknown type: 0, and with them it passes. Full unit suite is green on my side too.

Three things I'd like addressed before merging, in the inline comments:

1. onParserError can fire after the connection is already in Final (reproduced) and then emits a spurious "No event 'socketError' in state 'Final'" plus a second error after close(). Needs a Final guard or listener removal on state exit.
2. The once(tokenStreamParser, 'end') / abort races in the login paths leave an unobserved promise that the new 'error' emit can reject.
3. Request listeners are left attached on the failed request.

One more that is outside the diff but is the same bug one state over: SentAttention (around connection.ts:3880) does await once(tokenStreamParser, 'end') on its own parser, and the surrounding IIFE's .catch rethrows via process.nextTick. With Parser now re-emitting 'error', a malformed token in the attention-ack response rejects that once() and crashes the process, and the request callback is never invoked. Since this PR's stated goal is "surface token parse failures as request errors", I think that path should get the same socketError routing, either here or in a follow-up you're happy to own.

Not blocking: the copied buildLoginAckToken / drainMessage helpers in the test match what the sibling tests do, so no objection there.

---

Generated by Claude Code

[james-bg]

Thanks for the thorough review — all three inline items and the SentAttention case are addressed in 7ad2b18.

• The Final guard now lives in _onSocketError, as you suggested, and the two adjacent readMessage() catches plus both parser-error handlers route through it, so the dispatch/emit sequence is in one place.
• SentClientRequest detaches the request's cancel/pause/resume listeners before failing it. Your read was exactly right: without that, a late cancel() armed a cancel timer on the closed connection and its expiry hit the same Final hole.
• The four login/initial-SQL race sites observe their once() promise (same idiom as withAbortRace). No test for this one — the abort tears the socket down before a late parse error can arrive, so I couldn't construct it deterministically; happy to add one if you see a way.
• SentAttention wraps its once() and routes a rejection the same way, so a malformed attention acknowledgement fails the request instead of crashing.

New tests cover the close-from-handler, late-cancel and malformed-ack cases; each fails on the previous commit in the way you described (the first reproduces your No event 'socketError' in state 'Final' verbatim).

[arthurschreiber]

Thanks, this addresses everything. I re-verified on 7ad2b18:

• The three new tests each fail against the previous commit's connection.ts in the way described, and pass on this head.
• Lint and tsc are clean, the unit suite is green here, and it stays green on a trial merge with current master (the branch is behind #1773, but merges cleanly).
• I looked for a case where the Final guard in _onSocketError could hide a socket error that some path still needs and did not find one: every earlier state has its own socketError handler, and by Final the connection is already torn down.

On the untested abort-then-parse-error race: I agree it isn't deterministically constructible, since the abort destroys the socket before a late message can arrive. The no-op catch is the right defensive measure without a test.

LGTM.

---

Generated by Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 76.92308% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.90%. Comparing base (170fabc) to head (7ad2b18).
⚠️ Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/connection.ts	75.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1775      +/-   ##
==========================================
- Coverage   81.02%   80.90%   -0.12%     
==========================================
  Files          92       92              
  Lines        4948     4950       +2     
  Branches      938      933       -5     
==========================================
- Hits         4009     4005       -4     
- Misses        640      657      +17     
+ Partials      299      288      -11     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[github-actions]

🎉 This PR is included in version 20.2.2 🎉

The release is available on:

• npm package (@next dist-tag)
• GitHub release

Your semantic-release bot 📦🚀