Fix #1472: bound the mailbox escalation-toast dedupe set

[mohidmakhdoomi]

Summary

The VSCode mailbox escalation toast deduped "already shown" escalations with a seen Set that only ever grew — one permanent entry per escalated mailbox row for the lifetime of the extension host. It is now bounded, evicting on the mailbox row leaving the escalated set.

Fixes #1472

Root Cause

apps/vscode/src/notifications/mailbox-escalation-toast.ts — const seen = new Set<string>() is activation-scoped, and seen.add(payload.mailboxId) was its only mutation: no delete, no clear, anywhere in the module. Add-only container + extension-host lifetime = unbounded growth. Negligible per entry, but it never drains.

Fix

Evict on the mailbox row leaving the escalated set — the same signal that would let a legitimate re-escalation re-notify — mirroring how activateGateToasts prunes a builder that leaves the blocked set:

• The toast now takes the OverviewCache (already in scope at the single call site in extension.ts) and subscribes to onDidChange. OverviewData carries no per-row mailbox ids, only the workspace-level mailboxEscalated flag, so false — nothing held in this workspace is escalated — is the finest-grained "left the escalated set" signal available client-side. At that point every id in the set has left it, so the set is dropped whole. A null/absent cache read says nothing about the escalated set and never prunes.
• MAX_SEEN = 500 with oldest-first eviction (a Set iterates in insertion order) backstops the pathological window that stays escalated all day and so never observes that false.

No protocol or server change; the module and its call site are the whole surface.

Two notes checked against the server rather than assumed:

• db/mailbox.ts sets escalated = 1 once, guarded by escalated = 0, and never resets it; terminal rows are pruned. So the same mailboxId cannot legitimately re-escalate — eviction here is purely about memory, never about suppressing a notification.
• The prune cannot fire on a stale snapshot: OverviewCache.refresh() is last-write-wins by sequence, and the escalation SSE event itself triggers a refresh whose request starts after Tower flagged the row, so an older in-flight mailboxEscalated: false response can never commit after it.

Test Plan

• Regression test added — 4 new cases in src/__tests__/mailbox-escalation-toast.test.ts: an id re-toasts after de-escalation; the cap evicts oldest-first while the workspace never de-escalates; no prune while still escalated; no prune on an empty cache. Verified failing without the fix (2 failed / 10 passed with the eviction neutered) and passing with it (12/12).
• Build passes — pnpm check-types clean; porch build check green.
• All tests pass — 72 files / 850 tests in the vscode unit suite. pnpm lint reports one pre-existing warning in the unrelated src/commands/tunnel.ts.

🤖 Generated with Claude Code

[mohidmakhdoomi]

Architect integration review (risk tier: Low — isolated VSCode module)

Verdict: APPROVE (pending maintainer review — we are not maintainers; maintainer merges.)

Verified directly against the diff:

• Root cause is real and the fix is complete. Pre-PR seen Set at mailbox-escalation-toast.ts:30 had zero delete/clear paths; this PR adds both the semantic eviction (clear on mailboxEscalated: false via OverviewCache.onDidChange) and a MAX_SEEN=500 oldest-first backstop for a continuously-escalated window.
• Eviction key matches the tracking issue's guidance (Tracking: mailbox-first send follow-ups — gate correctness, interrupt semantics, attribution, visibility (post-1313) #1483 workstream D): the row leaving the escalated set is the same signal that would legitimize a re-notification. Mirrors the gate-toast.ts pruning precedent, so the two toast modules now age their dedupe state the same way.
• Server-side one-shot escalation semantics (escalated=1 guarded by escalated=0, never reset) means whole-set clearing on the workspace-level flag cannot cause duplicate toasts — verified the builder's analysis against db/mailbox.ts.
• Stale-snapshot race is closed by OverviewCache last-write-wins sequencing; the no-data guard correctly treats an empty cache as no signal.
• One-argument signature change, one call site (extension.ts:1522); four targeted regression tests cover both eviction paths, the still-escalated case, and the empty-cache guard.
• Interlock with Serializer convergence: route mailbox write edge through submitToSession (serialize gated deliveries vs interrupt/escape) #1365 (workstream B): that issue concerns write-path locking, not escalation semantics — no coupling risk to this eviction key.

Builder CMAP: gemini/codex/claude all APPROVE. No findings to address.

Parked for maintainer approval + merge.