Add a native cross-phase evolution orchestrator

[jramos]

Summary

Sequences the per-subsystem evolvers (skills → tools → prompts → code) from one YAML run-spec instead of invoking each by hand:

python -m evolution.orchestrator --spec examples/orchestrator/sample_run.yaml --dry-run
python -m evolution.orchestrator --spec examples/orchestrator/sample_run.yaml --only skills --only tools

Each phase runs as an isolated subprocess (fault containment, and it handles the CLI-only code phase uniformly). The orchestrator captures each phase's existing gate_decision.json at a deterministic --output-dir and writes a JSONL run history + summary.json/summary.md to its run root.

New package evolution/orchestrator/ (mirrors evolution/monitor/):

• spec.py — YAML run-spec loader/validator (SpecError at load, before any launch).
• adapters.py — one adapter per phase: argv builder + deterministic output dir + required fields + create-pr/cost-flag spelling.
• run.py — the sequencer with an injected phase_runner seam; gate-file-first reconcile.
• history.py — append-only run_history.jsonl + summary rendering (the deployable list is the human-in-loop handoff).
• __main__.py — the click CLI.

Propose-only by construction

• create_pr defaults off; the CLI strips every per-phase create_pr unless --allow-pr is passed, and a surviving PR phase must carry its spend ceiling (mirrors the monitor's --attempt-top guard). The orchestrator has no deploy/PR code path of its own.
• Verdict status is grounded in the gate file, never the evolvers' inconsistent exit codes (skills/tools exit 1 on setup errors before any gate; prompts exits 0 on dry-run). A clean run whose gate rejected the candidate is status=passed, decision=reject — not a failure. Only {failed, aborted} halt under --stop-on-error (default is continue-on-error; phases are independent).
• --resume skips phases already in the ledger; --only filters; --dry-run records the resolved argv without spawning.

Evolver change

Adds the missing --output-dir CLI option to the skills and tools evolvers (their evolve() already accepted output_dir; only the CLI wiring was absent) so the orchestrator captures every phase's run dir uniformly. Prompts and code already exposed it.

Why native (resolves our triage item for NousResearch#133)

The upstream PR is a real cross-phase orchestrator, but it couples phases in-process, targets a phase set this fork has diverged from (it has params — dropped here as a null axis — and lacks our code phase), and bundles unrelated changes (a Jaccard semantic-preservation constraint, a model-alias map, a judge cache, length penalties). This implementation borrows the shape (JSONL history, fault isolation, dry-run/phase-subset CLI) and rejects the in-process coupling, the bundled cruft, and the cross-run auto-scheduler (which would cross the propose-only boundary). docs/upstream_pr_triage.md is updated in this PR (row + review log + snapshot).

Test plan

• uv run pytest tests/orchestrator/ -q — 31 passed (spec/adapters/history/run/cli, fake-phase_runner seam, no LM spend)
• uv run pytest tests/skills/test_evolve_skill_output_dir.py tests/tools/test_evolve_tool_output_dir.py -q — 4 passed
• uv run pytest tests/ -q -m 'not slow' — 1746 passed, 4 deselected
• End-to-end --dry-run against the sample spec resolves correct, runnable argv for all four phases (every flag validated against each phase CLI's accepted options)
• pre-commit hygiene hooks pass