feat: add experiment registry and flight experiment verbs

[lewisnsmith]

Summary

• New src/experiments.ts registry module with race-safe file-per-experiment storage at ~/.flight/experiments/<name>.json
• New flight experiment {new,list,show,diff,export} command group for managing experiments and cross-run analysis
• Auto-registers unknown experiments on flight run --experiment with a one-line stderr hint

What this PR adds

src/experiments.ts

• ExperimentEntry type with name, created_at, tags, optional description, baseline_run_id, model_config, notes
• ensureExperimentRegistered — race-safe via { flag: "wx" } (O_EXCL); exactly one concurrent caller wins created: true
• createOrUpdateExperiment — merges patch (arrays replace, not append); preserves created_at
• getExperiment, listExperiments — graceful reads; invalid files skipped with a log warning
• experimentsDir — configurable via opts.dir for test isolation

src/cli.ts additions

• experiment new <name> — registers/updates an experiment
• experiment list — table with run counts from SQLite
• experiment show <name> — metadata + recent runs
• experiment diff <a> <b> — delegates to compareCommand with sessions from both experiments
• experiment export <name> — streams research JSONL per run to stdout (unbuffered)
• flight run --experiment now calls ensureExperimentRegistered and prints a hint on first use

Tests

• test/experiments.test.ts — 16 unit tests covering race safety, idempotency, merge semantics, invalid-file skipping
• test/cli-rename.test.ts — extended with 3 CLI-spawn tests (self-skip when tsx unavailable in worktrees)

Docs

• README, ARCHITECTURE, CLAUDE.md: added Experiment Registry section with schema, workflow, and CLI reference
• CHANGELOG: Added entry under 1.5.0

Note on pre-existing test failures

test/integration.test.ts and test/benchmark.test.ts fail in this worktree with ENOENT: spawn tsx — the worktree has no local node_modules. These failures pre-exist on main (verified). All 23 tests introduced in this PR pass.

Dependencies

This branch includes the PR 1 changes (feat/cli-logs-rename) bundled together. Should be rebased/merged after the feat/cli-logs-rename PR lands (or this can be rebased onto it).

Test plan

• npm run typecheck passes (exit 0)
• npm run lint passes (exit 0)
• npx vitest run test/experiments.test.ts — 16 tests pass
• flight experiment new smoke --description "hi" --tags a,b creates ~/.flight/experiments/smoke.json
• flight experiment list shows registry
• flight run --agent x --experiment new-one auto-creates + prints hint
• flight experiment diff a b exits 0 with a "no runs" message for zero-run case