The request says 'then update the composite, the TypeScript module and the widget'.

The widget source was last changed in the v1 UI job and contains no reference to the new Output.v2 (zP, zR, widthPct, active A/B/C events) or Output.subSignals fields; neither does src/widget/render.ts nor web/App.tsx. summary-v2.md admits this under Limits ('it does not render them yet'), so it is not overstated, but the deliverable is incomplete and the composite-settings width gate is invisible to a widget user.

Every v2 config, single-feature and composite, is scored with the accepted scorer's fixed bracket (1 ATR stop, 1.5 ATR target, 30-bar time exit) and the signal-to-signal exit family.

The request explicitly asked for ATR-stop widths, time stops, re-entry inside the band as confirmation, and a 15m-state filter for 1m/5m because 'the edge was in hit rate, not average size', and for family C to be tested as a gate for sizing and stop width. search_v2.py has no exit, stop or sizing parameters; the only 15m filter is the existing mtf='align' gate.

The summary states this under Limits, so the claims are not overstated, but the central question of the request (can exits turn the hit-rate lean into net edge) is unanswered. The hit rates reported (15m c_m7030 62%, 15m a_div 56%) with negative avg R are consistent with the owner's hypothesis and leave the sweep as necessary work.

Non-confirmation is decided from the bar distance to the last RSI run START of the same side, not from the RSI breach state on the price-touch bar. If the RSI breach began 4+ bars before the price outer-band touch and is still active, the touch is labelled 'price breach without an RSI breach' although RSI is beyond the 2.2 SD band on that bar, and b_both does not fire.

The TS port (featuresV2.ts line 55) has the same logic, so parity holds but both are wrong against the definition.

Fix: require rsi_b[i] == 0 (or no breach in the last 3 bars by state) for b_noconf, and let b_both accept an ongoing breach.

srs is built from lab/runs_v2 only; the 252 v1 attempts are counted in n_trials but their Sharpe ratios are not in the variance. The two inputs of sr0 describe different trial sets. Minor next to the variance problem above, but it should be fixed with it: load the v1 daily series (lab/runs/*.json) or state that the variance is v2-only.

The v1 cache ends 2026-11-12 00:00 UTC and the v2 fetch ran at 07:30 UTC the same day, so the 'new ETH' window is 7.5 hours (30 15m bars). No finalist fired; eth_new n = 0 for all four TFs. The summary says this plainly.

It means the only informative out-of-sample check is SOL, and the frozen configs have never been seen on ETH data outside their 160-day development set (which includes the burned v1 holdout). The check should be re-run once days, not hours, of new ETH bars exist; lab/search_v2.py final refuses to re-run while lab/runs_v2/final_checks.json exists.

mixes() builds 40 configs per TF around the v1 finalist parameters and re-scores them on 160 days that include the former holdout. The v1 finalists' holdout results (all negative) were known when v2 was designed, so the 160-day selection score is in-sample with respect to that information. The authors disclose this ('v1 holdout is burned ... treat it as development').

No quantitative residual bias is claimed in the artifacts, which is correct; readers should treat the dev selection scores as in-sample.

The claim that the width 'mid' gate improved the 15m v1 finalist cites 0.0107 vs 0.0064 for not_wide. The ungated v1 15m finalist (width none) was not part of the v2 search. Re-scored on the same 160 days it gives selection score 0.0054175 with n = 43 (net/day 0.00282, hit 0.561), so the claim holds and is actually conservative; the proper baseline should be stated.

Now: Placement only chooses above or below the whole 150 px widget. When neither side has enough space, the detail panel extends outside the viewport and clips its title and score.

Change to: Clamp the measured detail rectangle vertically within an 8 px viewport inset, allowing it to overlap the widget when needed; use a scrollable maximum height if the panel exceeds the viewport. Keep the existing focus, Escape and outside-tap behavior.

Why: The phone widget must expose readable per-timeframe details and sub-signals. Choosing the larger of two insufficient spaces does not keep those details visible.