Improve core dumps: per-op sync steps, structured failures, and failure-only LLM/test traces

[vishalramvelu]

Problem

When a PDD command fails, the debug snapshot saved to .pdd/core_dumps/ often lacks enough information to diagnose what went wrong. Example: pdd sync summarize_directory fails in a “5 consecutive fix operations” loop but the core dump may show:

• errors: [] even though the run failed
• steps has one opaque entry with missing/blank model
• no per-attempt breakdown of sync operations
• no test output for failed attempts
• no LLM prompt/response context for “LLM keeps generating broken code” scenarios
• terminal output contains ANSI noise

Goals

Make core dumps self-contained and actionable for debugging without reproduction, while keeping size reasonable and avoiding secret leakage.

Requirements

• Core dump payload
  Bump/maintain schema_version and include:
  argv, cwd, platform, pdd_version, timestamp_utc
  global_options, invoked_subcommands, total_cost
  steps (top-level commands), errors (structured), file_contents, terminal_output
• Top-level steps
  Ensure steps[*].model is never blank (default "unknown")
  If invoked_subcommands count exceeds returned results, still emit a step entry per command
• Terminal output
  Strip ANSI escapes (CSI + OSC) before saving terminal_output
• Sync operation visibility
  Auto-include .pdd/meta/_sync.log (JSONL) and .pdd/meta/_run.json in file_contents (size-capped)
  Ensure operation logs include duration so we can identify slow steps
  Expanded sync steps in core dump
  Add sync_steps derived from *_sync.log:
  one entry per operation row (skip type="event" rows)
  include: operation, success, cost (from actual_cost), model, duration, failure_summary, test_output_excerpt, source_log
• Test output excerpts (failed attempts)
  For failed test runs (esp. those used to drive fix), capture combined stdout/stderr
  Store as details.test_output_excerpt on the relevant operation log entry
  Truncate to ~5KB with an explicit truncation marker
• Structured logical failures (non-exception)
  If sync/fix loops terminate due to max retries, budget exhaustion, cycle detection, consecutive-op breaker, etc., record a structured entry in core dump errors:
  { command, type, message, details{ basename, language, counters/state } }
• LLM trace (moderate, failure-only)
  For each failed operation, record the final prompt/messages sent and the raw LLM response
  Keep it bounded:
  only the last pair per operation
  only persisted on failure
  truncate prompt/response to reasonable limits (e.g. <= 20k chars each)
  redact obvious secrets/tokens
  Store on the operation log entry as details.llm_trace = { prompt, response, model }

Acceptance Tests

Add/extend unit tests to verify:

• core dump has required fields + schema_version
• steps[*].model never blank
• ANSI stripping removes CSI+OSC
• file_contents includes *_sync.log and *_run.json when present
• sync_steps is populated from *_sync.log and includes source_log
• logical failures generate structured errors[]
• failed fix test runs attach truncated details.test_output_excerpt
• failed operations attach details.llm_trace

Context

Related: #680 (Core Dump Fixes), #230 (add LLM model CSV to core dump), #391 (confusing error output in core dumps)

[vishalramvelu]

Vishal target 3/19

[gltanaka]

Merged from #680 (closing as duplicate)

Two improvement suggestions from #680 not explicitly covered in this spec:

1. Include generated code diffs — show what changed between fix attempts (useful for diagnosing "LLM keeps regenerating similar broken code" patterns)
2. Include .pddrc and llm_model.csv configs in file_contents (if not already captured for the specific run) — these affect model selection and path resolution, both common failure causes

These complement the existing requirements here and should be considered during implementation.