feat: Surface details for chunk analysis and finding extraction failures

[dcramer]

Problem

When chunks fail to analyze or finding extractions fail, the user sees only an aggregate count in the summary:

⚠ 1 chunk failed to analyze
⚠ 10 finding extractions failed

There is no way to see which chunks or extractions failed, why they failed, or what the raw output was. The failure details exist internally (the SDK tracks them, Sentry gets breadcrumbs and metrics) but none of it is surfaced to the user.

This makes it impossible to tell whether failures are:

• Transient API issues (retries exhausted)
• Malformed skill output (extraction regex + LLM fallback both failed)
• Prompt too large for context window
• Something else entirely

Current State

What's tracked internally:

• analyzeHunk() logs to console.error() on failure (SDK returned no result, execution error, retries exhausted)
• parseHunkOutput() returns extractionFailed: true with extractionError and extractionPreview (first 200 chars of raw output)
• onExtractionFailure callback exists in analyze.ts but is not wired to CLI output in tasks.ts
• Sentry gets spans with hunk.failed, extraction.failed_count attributes

What the user sees: Just the count. No file paths, no error messages, no raw output.

Proposed Behavior

Verbose mode (-v)

Show per-failure details inline as they occur:

⚠ Chunk failed: src/utils/parser.ts:45-82 — SDK execution failed: context length exceeded
⚠ Extraction failed: src/cli/main.ts:120-155 — Could not parse findings from output

Debug mode (-vv)

Additionally show the raw output preview (first ~200 chars) and extraction error details for extraction failures.

Normal mode (default)

Keep current behavior (aggregate counts in summary) but add a hint:

⚠ 1 chunk failed to analyze (use -v for details)
⚠ 10 finding extractions failed (use -v for details)

Log file (optional, future)

Consider writing failure details to a debug log file (e.g. .warden/debug.log) that persists across runs, so users can inspect failures after the fact without needing to re-run in verbose mode.

Relevant Code

• Failure tracking: src/sdk/analyze.ts (lines 384-417 for hunk failures, 433-438 for extraction callback)
• Extraction logic: src/sdk/extract.ts (two-tier: regex then LLM fallback)
• Summary rendering: src/cli/output/reporter.ts (lines 210-231)
• Per-skill CI output: src/cli/terminal.ts (lines 199-203)
• Progress callbacks: src/cli/output/tasks.ts (onExtractionFailure callback exists but unused)