perf(task): reuse keys during journal union

[morluto]

Closes #77.

Compute journal union keys once per incoming/existing item, removing repeated prefix slices and key projections. Keep equality-based membership so unhashable values, deployed-environment identity and replay ordering remain supported.

Incoming deployed-environment items	Before	After	Speedup
100	0.189 ms	0.024 ms	7.7×
1,000	17.275 ms	1.797 ms	9.6×
3,000	151.949 ms	13.110 ms	11.6×

For 1,000 items merged into an empty list, key projections drop from 501,500 to 1,000. Membership comparisons remain quadratic. Timings are local union microbenchmarks, not complete task runs: macOS arm64, Python 3.12.13, baseline 602d730, median of five samples.

Reproduce:

.venv/bin/python tst/benchmarks/journal_union.py --base 602d730 --repeats 5

Validation:

• 63 focused journal tests passed; a new deterministic regression verifies projection counts and ordinary/move-to-end output for unhashable items.
• Full unit and protocol suites: 6,231 passed, 13 skipped.
• Benchmark compares production function bodies and checks equal outputs.
• Independent review completed with no actionable findings; plugin API check and diff checks passed.

This PR is independent of the task-graph, Explorer and existing context-diff PRs.

Confidence Score: 4/5

The PR appears safe to merge; the benchmark should also measure re-recording to show the cost of the changed move branch.

Findings

1.  Re-record speed goes unmeasured ▶

Fix with agent prompt

### Issue 1
tst/benchmarks/journal_union.py:71-75
The benchmark always passes an empty `existing` list and leaves `move_to_end` off. Re-recording a step uses the changed move branch, which now builds `existing_keys` for the existing items. Add a nonempty re-record case so maintainers can see whether that extra list affects its speed.

---

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

Summary

The task journal now computes union keys once per incoming and existing item, reducing repeated work while keeping equality matching and replay order. A benchmark and regression test cover the change.

• Task journal unions compute each item's key once.

_{Reviews (1) · Last reviewed commit: "perf(task): reuse journal union keys"}

[earakely-scale]

Taking a look