BLOCKING (shannon primary; carl light approved the model/conductor layer, which is nil-safe - the defect is in the view): app/views/releases/_deployment_average_chart.html.erb:25 - stages.dig(stage[:key], "average_seconds").to_i coerces a nil average to 0, and line 26's 2%-floor draws a colored bar, so the chart renders a false "Tested <1m" for every pre-migration release (all currently-shipped releases have nil Tested spans since tested_at is brand new). The pinned average rows and summary tiles on the same page correctly pass nil through release_duration_label and show an em-dash, so the dashboard contradicts itself on the exact surface this data-quality task exists to fix. FIX: keep nil (drop the .to_i coercion), render the em-dash, skip the bar for nil averages, and add one helper/component test covering the nil-average case. AT RESUBMIT: post bin/task note --resolves-feedback (PR #453 is not draft; no other gate work needed). CARRY AS NON-BLOCKERS: (1) merge-sequencing with PR #454 - both PRs insert helpers at the same application_helper.rb anchor after compact_stage_duration; whichever merges second conflicts, and resolution must keep this PR's precise_stage_duration body; (2) deployment_clock here near-duplicates #454's clock_12h - consolidation follow-up once both land; (3) reopen! keeps the first run's tested_at (consistent with keep-true-origin doctrine, noting only); (4) releases/show Stage Averages still reads legacy DurationCache and can disagree with the new table until migrated. Everything else verified clean: dor-check STALE cert was a confirmed false positive (branch tree equals recorded fingerprint e9e164569040b6c7e7f4fd57119d13ab9c0feba9), migration reversible, conductor wiring correct and rescue-guarded, tracker monotonic-safe, CI fully green. This is a one-line-plus-test rework; resubmit should be quick.

BLOCKING (shannon primary; jasper light approved full-page paths): the live stream sources do not follow in-frame filter changes. app/views/agents/activities.html.erb:21 renders agents/activities_live_streams OUTSIDE turbo-frame#aa-activities-frame (frame opens ~line 37), while every filter control (sidebar session toggles, active-filter chips, Clear all, pager) is an in-frame link_to. Frame navigation swaps only frame children, so the initial-load turbo-cable-stream-source tags persist: (1) load unfiltered then click a session filter -> page stays subscribed to the global agents_activities stream and nonmatching payloads still append to the filtered table (acceptance 2 fails); (2) load ?sessions=sess-A then Clear all -> page stays session-only and the unfiltered table misses other sessions' live appends (acceptance 3 fails). The fix works only on full page loads, which is why current tests pass - they assert full-page render output. FIX: move the activities_live_streams render inside the top of the turbo-frame so stream sources swap with filtered content (Turbo connects/disconnects cable sources on insert/remove), and add a test asserting the sources render within turbo-frame#aa-activities-frame. AT RESUBMIT: run gh pr ready 452 (PR is still draft) and post bin/task note --resolves-feedback. NON-BLOCKING nits: empty filtered page renders hb-empty with no thead#aa-activities-head target so the first matching live activity cannot append (pre-existing, fix optional); extra agent_activity PK lookup per action broadcast is acceptable. What is good: dual-broadcast design (global + session streams, single-scope subscription) is sound, and green CI on head a34c0834 supersedes the local SIGTERM full-suite bypass - keep it.