perf(task): prune validation and implicit scheduler graphs

[morluto]

Closes #76.

Prune unused task validation work and replace dense graphs with predecessor chains for wholly implicit sequential tasks. Mixed/explicit graphs retain their existing behavior. Both initial scheduling and retry rearming use the reduced representation.

Workload	Before	After	Speedup
Validate 3,000 implicit steps	97.93 ms	0.243 ms	403×
Validate 500 self-retrying steps	737.17 ms	0.049 ms	~15,000×
Build 3,000-step scheduler graph	220.57 ms	0.621 ms	355×

The graph uses 2,999 edges instead of 4,498,500. Peak traced allocations: 39.1 MB → 0.61 MB. Measurements are validation/graph microbenchmarks on macOS arm64, Python 3.12.13, against 602d730, with five timing samples. They do not measure complete task execution; earlier-target retries and mixed graphs retain their existing validation costs.

Reproduce:

PYTHONPATH=src:packages/agentenv-protocol/src .venv/bin/python tst/benchmarks/task_graphs.py --base 602d730 --steps 1000 3000 --retry-steps 500 --repeats 5

Validation:

• 157 focused task/retry tests passed, including actual execution, persisted retries, resume/end boundaries, callbacks and cancellation.
• Full unit and protocol suites: 6,237 passed, 13 skipped.
• Deterministic regressions check graph edge counts and absence of unused prefix slicing/self-retry ancestry.
• Independent review completed with no actionable findings; plugin API check and diff checks passed.

Two commits separate validation pruning from scheduler graph reduction. This PR is independent of the journal, Explorer and existing context-diff PRs.

Confidence Score: 4/5

The PR is not ready to merge because the documented graph benchmark fails against its baseline.

Findings

1.  Benchmark graph comparison crashes ▶
2.  Baseline uses newer helpers ▶

Fix with agent prompt

### Issue 1
src/agent_env/task/task.py:289
The new required `dependencies` field breaks the documented comparison with `main`. The benchmark gives the old `_build_scheduler_state` method the current `_SchedulerState` class, but the old method does not pass `dependencies`. The “scheduler graph” row raises `TypeError` instead of printing a measurement. Load the baseline state class with the baseline method, or adapt its return value.

### Issue 2
tst/benchmarks/task_graphs.py:51-56
If `--base` points to a revision with different dependency helpers, the benchmark runs that revision’s methods with helpers from the working tree. The printed speedup may then compare a mix of versions. Load the helpers from the selected revision or reject the comparison when they differ.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Task validation skips work it does not need, and wholly implicit task graphs use a predecessor chain with the same sequential run order. Mixed and explicit graphs keep their existing dependency behavior.

• Creating implicit tasks skips unused validation work.
• Implicit task steps follow a smaller predecessor chain.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Implicit step A] --> B[Implicit step B] --> C[Implicit step C]
  C --> D[Retry rearm counts stored predecessors]

Loading

_{Reviews (2) · Last reviewed commit: "refactor(task): build the scheduler's de..."}

[greptile-apps]

Baseline uses newer helpers

If --base points to a revision with different dependency helpers, the benchmark runs that revision’s methods with helpers from the working tree. The printed speedup may then compare a mix of versions. Load the helpers from the selected revision or reject the comparison when they differ.

Prompt To Fix With AI

This is a comment left during a code review.
Path: tst/benchmarks/task_graphs.py
Line: 51-56

Comment:
**Baseline uses newer helpers**

If `--base` points to a revision with different dependency helpers, the benchmark runs that revision’s methods with helpers from the working tree. The printed speedup may then compare a mix of versions. Load the helpers from the selected revision or reject the comparison when they differ.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Benchmark graph comparison crashes

The new required dependencies field breaks the documented comparison with main. The benchmark gives the old _build_scheduler_state method the current _SchedulerState class, but the old method does not pass dependencies. The “scheduler graph” row raises TypeError instead of printing a measurement. Load the baseline state class with the baseline method, or adapt its return value.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/task/task.py
Line: 289

Comment:
**Benchmark graph comparison crashes**

The new required `dependencies` field breaks the documented comparison with `main`. The benchmark gives the old `_build_scheduler_state` method the current `_SchedulerState` class, but the old method does not pass `dependencies`. The “scheduler graph” row raises `TypeError` instead of printing a measurement. Load the baseline state class with the baseline method, or adapt its return value.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.