feat(headless): add experimental todo task-tool baseline

[Astro-Han]

Summary

Default-off Harbor/headless task-tool experiment for #317, built on the task-system wiring from #537.

This PR ships one experimental baseline: todo_write with turn-tail guidance/replay. It records minimal task-tool usage in cell output/WAL/A/B reports and wires MAKA_MAX_STEPS through the Harbor host ai-sdk path.

The production default is unchanged: no task/todo tool is exposed unless the experiment env flag is set.

Why

Refs #317

We need a baseline for whether common task/todo tool surfaces help long-running benchmark tasks. Local A/B slices and trace review showed that CRUD-lite and bare todo_write mostly add schema cost without reliably changing model behavior, while todo_write with explicit turn-tail guidance makes the model externalize and update a task plan in most runs.

The evidence is still not strong enough for production enablement, so this PR keeps the path experimental, default-off, and removable.

Scope

Changed:

• Adds a headless task-ledger experiment tool surface with only todo_write.
• Enables the tool only under MAKA_CONTEXT_TASK_TOOLS=on, with MAKA_CONTEXT_TASK_REPLAY_MAX_CHARS for replay/guidance bounds.
• Injects replay/guidance through the turn-tail prompt so the base system prompt hash stays stable.
• Records minimal taskToolSummary through Harbor cell output, fixed-prompt WAL events, and A/B summaries/reports. Per-cell summary stores only todoWriteCalls; activation and activated attempt ids are derived from todoWriteCalls > 0.
• Wires runtime-policy A/B arm-local context env through Harbor.
• Wires host-side MAKA_MAX_STEPS into the Harbor ai-sdk backend path.
• Adds focused tests for default-off behavior, todo_write registration, replay caps, safe rendering, Harbor env isolation, output/WAL/report summaries, runtime-policy env isolation, and max-steps parsing.

Not included:

• Production/default task-tool enablement.
• CRUD-lite shape, CRUD tools, CRUD env entrypoint, CRUD store methods, or CRUD summary fields. CRUD-lite evidence remains historical experiment artifact only.
• Active tool prune behavior changes; this PR assumes the default-on active-prune setting validated in feat(headless): validate active-prune 2048 #340.
• A full benchmark rollout or acceptance claim.
• UI changes.
• Docs, changelog, migrations, or breaking changes.
• Automatically deleting the default-off experiment path if later evidence is negative.

Verification

• npm run -w @maka/headless test passed on 2026-07-06: 726 tests, 725 pass, 0 fail, 1 skipped. The skipped test requires a clean execution checkout to prove prompt-repo safety ordering.
• git diff --check passed.
• Commit split check: review fixes are split by revert reason, including the minimal summary cleanup commit.
• Local experiment artifact: /tmp/maka-task-tools-final/task-tools-final-1783289721020/summary.md.
• Manual trace review compared runtimeEventsPath function-call sequences, model_stream_started tool schema metadata, and task/todo call timing for off/CRUD/bare-todo/guided-todo arms.

User-facing impact

None. This is a default-off headless/Harbor experiment. No UI, docs, changelog, migrations, or breaking changes.

Reviewer notes

Experiment evidence:

Primary evidence uses the default active-prune setting from #340. The treatment under test is still only task/todo tools off vs on. CRUD-lite and bare todo_write are historical shape-screening evidence and are not retained as production paths in this PR.

Shape	Status in this PR	Case	A pass	B pass	Delta	B activated	B calls	Cost delta	Trace missing
CRUD-lite	excluded	crud-vs-off-prune-on-highcap	5/8	4/8	-0.125	1/8	1	+65.3%	4
bare todo_write	excluded	todo-vs-off-prune-on-highcap	5/8	3/8	-0.25	2/8	6	+15.0%	0
todo_write + guidance	retained	todo-guidance-vs-off-prune-on-highcap	4/8	5/8	+0.125	7/8	21	+18.8%	4

Trace findings:

• Historical CRUD-lite mostly behaved like extra schema, not a workflow change. It raised visible tool schema from about 2723 chars to 4442 chars, but only activated in 1/8 attempts in both high-cap cases.
• Historical bare todo_write also mostly behaved like extra schema. It raised visible tool schema from about 2723 chars to 3545 chars, but activated in only 2/8 attempts and often appeared late.
• Guided todo_write changed the trace shape: activation reached 7/8, the median first task-tool call position was 1, and many long tasks called todo_write within about 2-3 seconds.
• In successful long-task traces such as mailman and financial-document-processor, the model repeatedly rewrote the todo list from in_progress to completed and used it as a process anchor through final verification.
• The effect was not uniformly beneficial. In the default active-prune slice, guided todo_write improved large-scale-text-editing from budget-exhausted/fail to pass and reduced non-task calls slightly overall, but financial-document-processor and mteb-leaderboard became longer or more expensive.
• The only score lift in the default active-prune slice came from large-scale-text-editing. So the trace supports better long-task behavior shape, not a stable score-improvement claim.
• Failure traces still show false completion risk: overfull-hbox marked todo items complete but failed verification.

Interpretation:

• This PR ships only todo_write plus turn-tail guidance as the experimental baseline.
• CRUD-lite is excluded because adoption stayed at 1/8 in both high-cap cases.
• Bare todo_write is excluded because adoption stayed at 2/8 and score was worse in both high-cap cases.
• Guided todo_write increased candidate cost by about 18.8% in the default active-prune slice, so it should stay experimental until a longer-task confirmatory slice proves value.
• Any follow-up evaluation should assume active prune stays on by default and keep the main comparison as baseline tools off vs the selected task/todo treatment on. Variant-vs-variant runs can explain mechanism, but they should not replace the baseline acceptance question.

Review focus:

• MAKA_CONTEXT_TASK_TOOLS=on should be the only task-tool entrypoint.
• When enabled, the extra tool surface should be exactly todo_write.
• Per-cell taskToolSummary should stay minimal and store only todoWriteCalls; detailed tool-name counts remain in the existing toolSummary.actualToolCallCounts contract.
• A/B task-tool activation should be derived from todoWriteCalls > 0, while enabled-but-zero-call and budget-exhausted cells still count in the observed attempt denominator.
• Task replay/tool-result text should stay bounded and rendered through the safe ledger renderer.

Risk / rollback:

• Risk is limited to explicit experiment runs, but enabled runs can increase model cost.
• Individual commits are split by revert reason. If the experiment is not useful, remove or revert the default-off task-ledger experiment and Harbor wiring without changing normal headless/Harbor behavior.

Checklist

• Scope matches the PR title and excludes unrelated changes.
• Verification lists commands/results.
• User-facing impact, docs, changelog, migrations, and breaking changes are noted.
• Risk, rollback, or review focus is called out for non-trivial changes.
• UI changes are not applicable.