New 09-learning-system category testing learning capabilities:
- confidence-decay (10): staleness detection, decay-aware recall
- confidence-scoring (10): 1-10 scores, adjustment, thresholds
- dedup-correction (10): key supersession, no ghost entries
- cross-project (10): shared learnings with consent/isolation
- learn-management (10): search, prune, export, stats
- fp-learning-loop (10): FP tracking, dismissal, re-evaluation

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- build.rs extracts ironclaw git SHA from Cargo.lock at compile time
- Auto-populate framework_version with resolved SHA (was always empty)
- --ironclaw-rev flag: patches Cargo.toml, rebuilds, re-execs with new binary
- Per-category breakdown in run.json (pass_rate, avg_score, cost, time per tag)
- Tags persisted on TaskResult for category aggregation
- Compare output shows per-category deltas and framework versions
- scripts/bench-ab.sh for end-to-end two-ref comparison
- Fix pre-existing test failures from ironclaw API changes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

New 10-real-world-workflows category with complex, multi-turn scenarios
requiring Python code execution and tool composition:

- financial-analysis (10): earnings analysis, portfolio rebalance, stock
  screening, revenue forecasting, expense anomaly detection, options
  payoff, invoice reconciliation, tax optimization
- blockchain-ops (10): tx parsing, wallet tracking, DeFi yields, gas
  optimization, NFT analysis, contract audit, bridge monitoring, MEV
  detection, DAO treasury
- sports-analytics (10): player trends, fantasy draft, game prediction,
  salary cap, injury impact, trade values, season simulation, draft
  picks, betting odds, coaching strategy
- long-context-stress (10): massive log triage, multi-doc synthesis,
  codebase review marathon (8 turns), meeting consolidation, compliance
  audit, incident postmortem, contract review, KB cleanup, resume
  screening with bias resistance, context overflow recovery
- data-engineering (10): CSV insights, API ETL, log parsing, data quality,
  time series, schema migration, fuzzy multi-source join, report
  generation, PII anonymization, streaming data simulator

Total ironclaw-v2 suite: 195 tasks, 528 turns across 10 categories.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Fix 10 tasks with response_matches as array (must be string)
- Fix duplicate response_matches key in data-quality-report.json
- build.rs detects engine_v2 field in AgentConfig via cargo git cache
- Sets cfg(ironclaw_engine_v2) so runner.rs conditionally includes it
- Register expected cfg in Cargo.toml lints to suppress warnings

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Clone state before move in set_state() logging
- Replace BenchError::Other (nonexistent) with BenchError::Config
- Fix Vec<str> type error in extract_error_patterns
- Prefix unused parameter with underscore
- Use explicit PathBuf type annotation for directory entry

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Suppress dead_code warnings on post_mortem_mission module
- Fix unused variable (_deployment_id)
- Apply clippy suggestion: map_or -> is_none_or
- Run cargo fmt across all source files

Build, clippy, and fmt are now all clean with zero warnings.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

CLI: `nearai-bench post-mortem <run-id>` analyzes failures from a
completed benchmark run. Outputs failure stage breakdown (scoring vs
timeout vs setup), per-category failure rates, and recommendations.
Supports --format json and --force for re-analysis. Results cached
as post_mortem.json alongside run.json.

Runner: `--auto-post-mortem` flag (or `auto_post_mortem = true` in
TOML config) auto-generates post_mortem.json after any run with
failures and prints the summary.

Bridge layer converts TaskResult failures into the PostMortemMission
event model, categorizing by failure stage (timeout, setup, execution,
scoring) and aggregating root causes across all failing tasks.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Enrich TaskResult with full data needed for analysis and training:

- Per-LLM-call detail: input/output tokens, duration, had_tool_calls
  (was only aggregate totals before)
- Tool call inputs/outputs: capture parameters and result previews
  (was only name + success boolean)
- Full conversation history: all user/assistant turns in order
  (was only final response)
- System prompt field: identity/SOUL.md content used for the task

All new fields use #[serde(default, skip_serializing_if)] for
backward compatibility with existing results.

Channel now tracks tool I/O flow: ToolStarted -> ToolResult (preview)
-> ToolCompleted (parameters/error), assembling the full picture.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Each task now runs with its own temp directory as the shell's working
directory. Agent-created files (Python scripts, data outputs, etc.)
land in the temp dir instead of polluting the repo root.

The runner creates a tempdir per task and registers a ShellTool with
with_working_dir() pointing to it, replacing the adapter's default
shell. The temp dir is cleaned up automatically when the task completes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Run d01ec670: 24.6% pass rate, 0.592 avg score, $4.74 cost, 86min
Includes post_mortem.json with failure analysis.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The codebase-review-marathon task had a well-known Stripe test key
(sk_live_4eC39HqLyjWD...) in its planted vulnerability data. Replaced
with an obviously fake key that won't trigger GitHub's secret scanner.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>