fix(watch): keep no-farmable grace alive across transient active-drop blips

[rxm96]

Symptom

The engine status panel could get stuck on "Watching, but target currently not watchable" for a very long time (observed: 1h17m) with SUPPRESSION none, COOLDOWNS none, still actively "Watching", and a frozen no-progress tracker showing 0 recovery attempts. A full app refresh cleared it — i.e. it was a wedged runtime state, not expected parking.

Root cause

Watchability is decided by drop state, not channels, so a target with only claimable/blocked/upcoming drops is correctly canWatchTarget = false. When that happens while watching, decideWatchingNoFarmable is supposed to wait out NO_FARMABLE_DROP_GRACE_MS (30s) and then give up: cooldown + retarget + stop + stall-stop. That pure logic is correct and already tested.

The bug is in the executor useStallRecovery: it cleared the no-farmable grace marker (noFarmableDropRef.current = null) on every pass through the no-progress recovery branch — i.e. whenever activeDropInfo was momentarily present. If the active-drop selection oscillates (e.g. an upcoming/unlinked drop surfacing for a tick while canWatchTarget stays false), the marker is reset on each blip, the 30s grace never completes, the give-up path never fires, and the stale stall tracker keeps showing 0 attempts.

This matches all observed symptoms: no give-up, no cooldown, no suppression, frozen tracker.

Fix

Preserve the grace marker across a transient activeDropInfo while the target stays unwatchable; clear it only once the target is genuinely watchable again (canWatchTarget === true). Extracted as the pure reconcileNoFarmableOnActiveDrop(marker, canWatchTarget) and wired into the executor in place of the unconditional reset.

Entry condition of the no-farmable branch is intentionally left on !activeDropInfo (not !canWatchTarget) so the existing channel-switch recovery — for the case where watchable drops exist but not on the current channel — is preserved.

Tests

TDD: added a failing regression first, then the fix.

• reconcileNoFarmableOnActiveDrop units: preserve while unwatchable, clear when watchable, null-safe.
• End-to-end flicker regression: seed grace → active-drop blip (marker must survive) → grace elapsed → give-up (stop-watching + resetStallTracking) fires.

npm test (524 passing), npm run typecheck (both tsconfigs), prettier + eslint on changed files: all clean.

Notes

Found via a user-reported screenshot. Workaround until merged: stop/restart watching (or toggle the master switch) re-initializes the refs.

🤖 Generated with Claude Code