daily-fix: instant wedge lane for Prompt-is-too-long

Overview / Motivation

Auto-filed by /daily 2026-07-16 from transcript mining (#1335, ebff95d1: context-ceiling wedge 12:48-13:53Z, no compact_boundary fired, 4 consecutive dead tick turns).

Goal

Cut context-ceiling wedge recovery from ~65+ min (3-wake accumulation) to one watcher tick.

Workflow gap

• Bug observed: a session whose turn dies with the synthetic 'Prompt is too long' API error is unrecoverable in-session by definition (every future turn fails identically), but the watcher's prompt-wedge lane waits for the generic >=3-failed-wake accumulation across 45-min tick spacing — #1335 lost ~65 min wedged (4 dead tick turns 12:48-13:50Z) with a one-line fix unapplied
• Why it is a workflow gap: The wedge lane exists to bound dead-session time; a deterministic, textually-identifiable unrecoverable failure should not wait for statistical accumulation.
• Confidence (emitter): medium
• verified-at-filing: grep -n 'Prompt is too long' scripts/autonomous_session_watch.py -> 0 hits (no subclass handling — absence claim); wedge-lane counters exist (EPM_TICK_WEDGE_MIN_FAILED_TURNS lineage, #1127/#1209/#1241)

Proposed change (candidate diff sketch — refine in planning)

treat 'Prompt is too long' as an instant prompt-wedge subclass: on the FIRST failed wake turn whose api-error text matches it, stop + force-respawn (existing episode-belt + day caps apply)

Scope / surfaces

• Primary target: scripts/autonomous_session_watch.py
• Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• fingerprint: 89073e126181