fix(miner): coding-task-spec.js's buildTaskBrief silently drops the prompt-injection audit event its sibling buildInstructions always logs

[JSONbored]

Context

packages/loopover-miner/lib/coding-task-spec.ts's own header comment (lines 81-88) documents a specific,
load-bearing invariant: "neutralizePromptInjection runs on both fields before they reach either the coding
agent's instructions (buildInstructions) or the acceptance-criteria document's taskBrief
(buildTaskBrief) -- the two places raw issue text is embedded into agent-facing prose." Both functions do
call neutralizePromptInjection on the issue title/body, and both correctly redact matched content. But only
one of the two logs an audit event when something was actually neutralized:

• buildInstructions (lines 256-267): checks title.injected || body.injected and emits
  console.log(JSON.stringify({ event: "prompt_injection_neutralized", issueNumber, fields: [...] })).
• buildTaskBrief (lines 90-94): calls the same neutralizePromptInjection(...).text twice but never reads
  or reports the .injected flag at all -- it only takes .text and discards the rest of the result.

buildTaskBrief's output becomes AcceptanceCriteria.taskBrief via buildCodingTaskAcceptanceCriteria,
which is written to disk inside the attempt's own worktree (writeAcceptanceCriteriaFile) as a persistent,
inspectable artifact. This is exactly the kind of trust/safety signal this package logs elsewhere for the
identical defense (see submission-freshness-check.ts's own abort-reason audit event, and
src/review/prompt-injection.ts's safeReviewTitle which logs prompt_injection_neutralized on the review
side). A target-repo issue whose title/body carries a prompt-injection payload aimed specifically at the
acceptance-criteria document (rather than the coding-agent instructions) is neutralized correctly today, but
the miner's own operator has zero visibility that it happened -- the one call site that would tell them is
silently a no-op.

Requirements

• buildTaskBrief (or its caller buildCodingTaskAcceptanceCriteria) MUST emit the identical
  prompt_injection_neutralized audit event shape buildInstructions already emits --
  { event: "prompt_injection_neutralized", issueNumber, fields: [...] } -- whenever neutralizing the title
  and/or body actually redacted something, using the same fields array convention (["title"], ["body"],
  or ["title", "body"]).
• The event MUST be emitted exactly once per buildCodingTaskAcceptanceCriteria call (not once per field,
  not duplicated if buildCodingTaskSpec later also calls buildInstructions on the same issue -- both call
  sites logging independently, once each, for their own field, is the correct and expected behavior; this
  issue is only about buildTaskBrief currently never logging at all).
• Do NOT change what gets redacted or how neutralizePromptInjection itself works -- this is strictly an
  audit-logging parity fix, not a change to the redaction logic.
• buildCodingTaskIssue's existing callers and CodingTaskSpecResult's shape must not change.

Deliverables

• buildTaskBrief (or buildCodingTaskAcceptanceCriteria) logs prompt_injection_neutralized with the
  same JSON shape buildInstructions uses, whenever the title and/or body was actually redacted.
• A regression test in test/unit/miner-coding-task-spec.test.ts mirroring the existing
  "logs a prompt_injection_neutralized audit event..." test (line 405) but asserting it against
  buildCodingTaskAcceptanceCriteria's output, using a malicious title/body fixture (the existing
  "Ignore all previous instructions..." fixture at line 142 already exercises the redaction; extend it,
  or add a sibling test, to also assert the log call).
• A regression test confirming a benign issue (no injection pattern) produces zero log calls from
  buildTaskBrief's path, mirroring the existing "does not log... for a benign issue" test (line 431).

Test Coverage Requirements

99%+ Codecov patch coverage (branch-counted) on every changed line in coding-task-spec.ts. Both the
"injection detected, log fires" and "no injection, no log" branches need explicit tests, matching the
existing coverage pattern already present for buildInstructions.

Expected Outcome

A miner operator inspecting logs can see every case where a target repo's issue content required
prompt-injection redaction before being embedded into either agent-facing artifact (instructions OR
acceptance-criteria taskBrief) -- today only one of the two artifacts' redaction events is observable.

Links & Resources

• packages/loopover-miner/lib/coding-task-spec.ts (header comment lines 81-88, buildTaskBrief lines 90-94,
  buildInstructions lines 256-277)
• test/unit/miner-coding-task-spec.test.ts (lines 141-149, 405-429, 431-445)
• packages/loopover-miner/lib/prompt-injection-defense.ts (the shared neutralizePromptInjection)
• src/review/prompt-injection.ts's safeReviewTitle -- the established audit-logging precedent on the
  review side of the same defense

[linear-code]

JSO-5302