PR #27979 runtime verification, Linux Boat VM
main: 30084f0018c5 (github/main at verification start)
PR: 626491b1d9f32c769ae2fe82fb88f26122b5e21c

Setup: separate main worktree, vp i in both checkouts, one background dev
server at a time, isolated empty T3 homes, no seeded database or mocks.
Project /home/user/main added through the live project CLI. Real Codex
provider gpt-6.1-sol completed "reader-pool-ok" on each server.

Controlled scheduling reproduction through authenticated server HTTP APIs:
Node inspector paused the four real SQLite reader workers. Four concurrent
GET /api/orchestration/shell calls occupied the four reader slots. Then
GET /api/projects queued. Only reader slot 2 was resumed. A later projects
request verified that this reader was available. Finally all were resumed.
No application code or database contents were altered by instrumentation.
The timestamps below show ordering under debugger control, not a latency
benchmark. Pauses were requested before the workers received each query.

$ timeout 20 node /tmp/saturate.mjs main 16733
main +1605ms queued completion before release: false
main +1606ms resumed reader slot 2; slots 1, 3, 4 remain paused
main +1612ms held shell 2: HTTP 200, 4880 bytes
main +2111ms later projects read: HTTP 200, 1401 bytes
main +3111ms queued completion with only slot 2 available: false
main +3115ms queued projects: HTTP 200, 1401 bytes
main +3116ms held shell 1: HTTP 200, 4880 bytes
main +3156ms held shell 3: HTTP 200, 4880 bytes
main +3200ms held shell 4: HTTP 200, 4880 bytes

$ timeout 20 node /tmp/saturate.mjs pr 13773
pr +1661ms queued completion before release: false
pr +1662ms resumed reader slot 2; slots 1, 3, 4 remain paused
pr +1674ms held shell 2: HTTP 200, 4884 bytes
pr +1674ms queued projects: HTTP 200, 1405 bytes
pr +2165ms later projects read: HTTP 200, 1405 bytes
pr +3165ms queued completion with only slot 2 available: true
pr +3171ms held shell 1: HTTP 200, 4884 bytes
pr +3210ms held shell 3: HTTP 200, 4884 bytes
pr +3251ms held shell 4: HTTP 200, 4884 bytes

Result: main completed a later read while the earlier queued read stayed
blocked until slot 1 resumed. PR completed the queued read using slot 2,
while slots 1, 3, and 4 remained paused. All held shell requests then
completed HTTP 200 after the remaining readers resumed.

Web client: Playwright from repo dependencies, /usr/bin/google-chrome.
Opened the real provider thread, submitted follow-up messages using the
composer, observed follow-up-ok and reader-reuse-ok from gpt-6.1-sol.
Reload restored both original and follow-up responses.

Blacksmith: all 20 existing nodeSqliteClient tests passed, including
first-free reader, queued transaction cancellation, lock-wait metric,
read-only enforcement, nested transactions and worker failure fallback.