Plan-path task_result always reports secs: 0.0

[wseaton]

Every plan-path task_result goes out with secs: 0.0. task_result_event (crucible/src/plan/cli.rs:362) hardcodes it because plan::exec::TaskResult (crucible/src/plan/exec.rs:141) carries status/attempts/cost_usd/output/note/fanout and no timing, and Attempt (exec.rs:47) is outcome + cost_usd. The only clock in the executor is the plan-wide wall_clock ceiling (exec.rs:274). The secs field on the wire event is populated by the measure-DAG emitter and defaulted by everyone else.

Downstream is faithful to the lie: the controller stores 0.0 in run_task_results.secs, and 0.0 is a number, so the run page renders 0.0s per task and every duration bar collapses to its floor. It cannot tell "took 0s" from "not measured" (it is patched for now to treat a run where nothing is positive as untimed).

Plan

Time at the runner boundary, not in the executor. Under run_many a batch runs concurrently, so the executor only sees batch wall time; the runner is the one thing that knows per-item elapsed.

1. Attempt gains secs: f64. Add a constructor that takes the Instant the runner started at so the ~50 construction sites (most are tests, or undispatched paths where 0.0 is honest) do not each open a stopwatch. The real runners are plan/runner.rs:26, plan/harness.rs:138,208, loop_graph.rs:447,452,705,747,1187,1224; each measures around the actual work. The trait-default run_many (exec.rs:81) inherits it by delegating to run.
2. exec::TaskResult gains secs: f64, summed across retries in run_with_retries (exec.rs:787) and per-index in run_batch_with_retries (exec.rs:883), the same way cost_usd accumulates. fold_instances (exec.rs:697) sums instance secs; undispatched and reduce_top_k stay 0.0.
3. task_result_event passes r.secs instead of 0.0. The three loop_graph.rs callers (133, 375, 886) and the plan run caller (cli.rs:686) need nothing.
4. SessionEvent::TaskResult.secs doc gets the plan-path semantics: wall clock across all attempts of that report, 0.0 only for a task that never dispatched.

Tests

• Executor: a TaskRunner test double that sleeps (or returns a fixed secs) proves run_with_retries sums across transport retries and run_batch_with_retries keeps per-item figures rather than the batch's wall time.
• task_result_event round-trip (cli.rs:1070) asserts secs survives encode/decode with a non-zero value.
• fold_instances sums instances.
• The undispatched paths (blocked, truncated) still report 0.0.

Rollout

Pin bump in crucible-domains (core-pin.toml + the crucible/crucible-contract/forge git revs in crucible-controller/Cargo.toml) and a loop image rebuild. The controller's ingest (ingest.rs) and run_task_results.secs already carry f64; no migration. Once the pin lands, drop the controller UI's "all zeros means untimed" normalization (ui/src/pages/taskGraph.ts timedResults).