feat(loop): agent-suggested parallel batches + end-of-run verify + loop graph DAG (hew-lf40)

[droidnoob]

Closes the hew-lf40 epic — agent-suggested parallel batches for hew loop run --jobs N, plus end-of-run test verification and a DAG renderer for loop iteration history.

Why this exists

DECISION:loop-parallel-overlap-policy ("trust the graph") shipped in v1 as a deliberate punt: bd dep edges encode safety, and conflicts get caught at merge-back time. That's correct for sparse graphs but bites when two independent tasks touch the same file. The 2026-05-29 autonomous run made the cost real — loop_log.rs overlap between hew-2cq and hew-6nxs required a manual rebase even though both branches were green in isolation.

This epic layers informed batching on top of trust-the-graph without contradicting it:

1. Iter agent's own suggestion — when the agent that just closed iter N emits a next_iteration: [task_ids] block, the dispatcher honors it for the next tick. Cheapest signal: the agent already has full context.
2. Dedicated planner runtime — between iters, when (1) is silent, a small claude -p / codex exec call (capped by loop.planner.budget_tokens, default 10k) reads the bd-ready set + recent symbol-touch sets and returns the next batch. Never truncates context to fit budget — skips cleanly instead.
3. Trust-the-graph floor — dispatch_tick intersects the batch with bd ready. The batch can only narrow the candidate set, never expand it. Floor is locked.

What lands (8 atomic commits)

commit	task	what
108d148	hew-58ac	hew_core::batch_plan module — BatchPlan + BatchSource enum + atomic file I/O at .hew/loop/<run-id>/batch-NNN.json, schema_version=1
e33abb0	hew-7klt	batch_plan_parse::extract_next_iteration — tolerant parser for the agent's close-output block (fenced next_iteration form + <next_iteration> XML form), filters malformed task IDs
f58ff12	hew-pxw9	spawn_planner — subprocess with pre-spawn budget check; failure modes all return BatchPlan::Skipped { reason } rather than propagating errors
48506d9	hew-rplg	Dispatcher threading: Dispatcher::new accepts Option<BatchPlan>; dispatch_tick filters by batch ∩ bd_ready; ready_seen reflects post-filter
31ef9ff	hew-7k1m	CLI + config — --no-planner, --planner-budget, loop.planner.* schema; iter-end hook chooses agent → planner → skipped
5e595fa	hew-z7rz	hew loop summary adds planner: agent=N, runtime=M, fallback=K; docs/LOOP.md "Batch planner" section; CHANGELOG entry; new DECISION:loop-batch-planner-floor memory
dbe56b4	hew-bon7	End-of-run verify step: stack-detected test command (Rust → Node → Python → Go → Make/Just), loop.end_of_run.verify_tests config + --verify-tests flag, opt-in default false, budget-capped, writes verify.log and STATUS:loop-verify-failed: memory on failure
42014ba	hew-m7lq	hew loop graph — DAG renderer over iter + batch + run logs. Outputs mermaid (default), dot, or ASCII. Handles incomplete iters, cancelled runs, runtime-error-with-empty-stderr, backpressure rollback, verify outcomes, parallel worker swimlanes

Bonus: the hew loop graph unhappy paths the user flagged

Per the chore body, the graph must render the cases where things didn't go cleanly:

case	node treatment
incomplete iter (started_at, no ended_at)	⋯ glyph, dashed border, partial label
cancelled mid-run (stop_reason: Cancelled)	⊘ glyph, gray, [Cancelled at <ts>] annotation
runtime_error with empty stderr (the 2h hang case)	✗ glyph, (no stderr — possibly hung) annotation
backpressure_fail with rollback	↺ edge back to previous iter's HEAD sha
verify failed	red verify node + top 3 failed test names

CI parity

• Local: cargo test --workspace — 41 suites green, 0 failed.
• bd-scrubbed parity check (mimicking CI's lack of bd): loop_scope_e2e 7/7, loop_backpressure 14/14. The precheck-before-discover guard from feat(loop): hew loop run --scope={ready|epics} flag (hew-b3yl) #58 still holds.

Backward compat

• Legacy runs (no batch-*.json files in run-dir): dispatcher reads read() == None, falls through to bd ready. Byte-identical to today.
• Legacy run.json (no verify_outcome field): #[serde(default)] → None, summary line omitted.
• --jobs=1 (the default): batch-plan layer skipped entirely. Single-worker loops don't write batch files.

Non-goals (v1)

• Static touches-overlap analysis (would parse description prose; brittle).
• Cross-run batch memory.
• Auto-fix on verify failure.
• Live-updating graph (websocket/fswatch).

🤖 Generated with Claude Code