Add deterministic abortable run-once Pi session streaming

[rochecompaan]

Summary

• Allocate exact parent Pi session files for observed run-once prompts and pass them with --session.
• Stream exact session JSONL serially with callback backpressure, strict errors, UTF-8-safe reads, cancellation, and aggregate terminal causes.
• Keep triage on legacy session-dir observation while adding regression coverage.

Validation

• node --test src/cli/commands/run-once/pi.test.ts src/cli/commands/triage/*.test.ts — 187 tests passed.
• npm run test:run-once — 431 tests passed.
• npm run lint:ts — passed with zero warnings.
• npm run lint:md — 177 files, zero errors.
• git diff -- package.json npm-shrinkwrap.json package-lock.json — no output.
• Final validation-readiness reviewer also ran npm test — 1,154 tests passed.

Reviews

• Final Codex full-worktree review passed after accepted fixes.
• Final thermo-nuclear review passed with one deferred minor test-organization suggestion.
• Final validation-readiness review passed all required commands.

Closes #123

Patchmill run cost

Stage	Model	Prompt tokens	Output tokens	Estimated cost (USD)
Planning	gpt-5.5	1,549,932	18,856	$1.7443
Development environment	gpt-5.5	670,781	2,262	$0.7353
Implementation	gpt-5.5	6,136,856	18,466	$4.5119
Implementation	gpt-5.6-sol	14,146,920	82,481	$14.2658
Implementation	gpt-5.6-terra	13,632,353	47,926	$6.3315
Implementation subtotal		33,916,129	148,873	$25.1092
Total		36,136,842	169,991	$27.5888

Prompt tokens include uncached input, cache reads, and cache writes. Cost is an estimate based on Pi's recorded model pricing and includes parent and subagent sessions.