fix(ci): repair the lint gate and the CRLF-fragile source scan on main

[fxogg06-jpg]

Why

main at 928f816 has a red verify / regression job. The failing step is 7. Run bun run lint, which means step 8 (bun run test) is skipped — so the regression suite has not actually run on this commit. Fixing only the lint error would turn CI green while the second defect below stays hidden; both are included here.

Reproduced locally on a clean checkout of 928f816.

1. bun run lint fails with 4x no-useless-assignment

src/controllers/anthropic.js
  1265:7  error  The value assigned to 'promptTokens' is not used in subsequent statements
  1266:7  error  The value assigned to 'completionTokens' is not used in subsequent statements
  2044:7  error  The value assigned to 'promptTokens' is not used in subsequent statements
  2045:7  error  The value assigned to 'completionTokens' is not used in subsequent statements
✖ 4 problems (4 errors, 0 warnings)

In both handleAnthropicStream and handleAnthropicNonStream, the two counters were initialised to 0 and then unconditionally overwritten from the authoritative usage object:

let promptTokens = 0;
let completionTokens = 0;
...
const usage = reportUsage(upstreamUsage, ...);
promptTokens = usage.prompt_tokens;
completionTokens = usage.completion_tokens;

The initialisers are dead stores: nothing reads the variables before the assignment. The locals are write-only intermediates, so they are removed and both values are taken as const directly from usage at the single point where they are resolved.

No behaviour change — the values passed to attributeChatUsage and written into the emitted usage payload are byte-for-byte the same. This is a pure dead-store removal, not a change to how usage is resolved or estimated.

2. tests/tool-prompt.test.js:805 fails on any CRLF checkout

✖ lockstep: ningun sitio de prompt/hint re-ensena la forma nativa <tool_call>
  AssertionError: ../src/controllers/anthropic.js: cadena con la forma nativa
  <tool_call> legible por el modelo

That test scans chat.js / anthropic.js / request.js to pin at source level that no model-readable string re-teaches the native <tool_call> form. It deliberately strips comments first, because the rationale comments in the tree legitimately name <tool_call>:

const code = src
  .replace(/\/\*[\s\S]*?\*\//g, '')
  .split('\n')
  .map(line => line.replace(/\/\/.*$/, ''))   // <-- the bug
  .join('\n')

The per-line strip is anchored with $. On CRLF input . does not match \r, so $ matches before the CR, and because that regex has no g flag the engine backtracks to the empty match at position 0 — leaving the entire comment body in code, including the <tool_call> the test means to exclude. It then trips on its own rationale comments:

// 错误列表（hasParseError 永远为真，即使重试本身成功），而一个被截断的 <tool_call>

Bounding the strip by the newline instead is what the test intends, and behaves identically on LF:

.map(line => line.replace(/\/\/[^\r\n]*/, ''))

Also worth noting for triage: this is a git checkout dependency, not a platform one — core.autocrlf=true (default on Windows) is enough to trigger it, and the repo has no .gitattributes pinning * text=auto eol=lf. Adding one would be a more structural follow-up, but I kept this PR to the two defects so it stays reviewable.

Verification

Run on Windows, Bun 1.4.2 (from package.json), Node v24.21.0, matching the workflow's environment:

Gate	Before	After
bun run lint	exit 1, 4 errors	exit 0
bun run test	never runs in CI (skipped after lint)	PASS — 1146 tests / 134 suites / 0 fail
bun run build:frontend	n/a	exit 0
bun run test:bun	n/a	PASS (8/8)

1146 / 134 / 0 fail matches tests/expected-counts.json exactly, so the count gate is satisfied rather than merely non-failing. The smoke run covers real Bun startup, login, frontend, WASM, HTTP/SSE (including a 16 s upstream delay), and agent mid-stream quota failover.

Verified the patch applies cleanly onto a fresh clone of main at 928f816 before opening this PR.