Make main green: source-form e2e harness, section-mode launch reads, checks-job repairs

[TheGreatAxios]

Main's CI has been red since the milestone merge (#108) — both jobs. This
puts them back green and fixes the real regressions the red was hiding.

walking-skeleton

The e2e harness still spoke the pre-cutover deploy dialect. It pushed
a bare workflow.json into the asset repo and deployed with
{assetId, sources}. The source-form gate rejects the envelope
(path-violation: … the workflow.json envelope form is no longer supported), so seven suites died at the push. The harness now publishes
through the platform's own createGitWorkflowPusher (rendering
@corbits/workflow-source's tree) and carries the pushed commit into a
{source, entry} deploy body — the same path the seed and the product
take, so the suite can't drift from them again.

Agent-chat reuse compared definition row ids. A code-sourced deploy
projects a fresh workflow_definition row per frozen wire projection, so
the id a chat recorded at creation and the id the picker offers later are
routinely different rows over the one asset that is the agent:
"reopen my chat with this agent" minted a duplicate instead. Reuse now
compares the asset.

Fixture alignment alongside those: the folded-run backfill insert
predated workbench_launch's NOT NULL current_run_id; local-rip's
planner gate counted definition rows (re-projected per run) and now names
the agents that appeared, which is what the gate was always about; and
the chat settings test ran on bun's 5s default while its siblings allow
90s.

(The section-mode launch read — readFoldedBody choking on CL-6329's
onTrigger turn — was fixed on main in parallel while this branch was in
flight; this branch keeps main's implementation and drops its own.)

checks

Lint failed on prettier drift, which meant eslint and the unit suite
never ran — and both were red behind it. Fixed: the formatting, three
unused symbols left by the section swap, the task-planner fake db (which
predated the launch path's frozen-projection read), and agent_turns'
missing check:no-product-tenancy allowlist ruling.

The toast suite's order-dependent failure is fixed at the cause.
mock.module rewrites the module registry for the whole process and bun
offers no way to take it back, so three suites that stubbed
@corbits/react-ui's toast left the stub installed for every file
loaded after them; toast-single-system.test.tsx renders the real
toaster and observed nothing. The shared spy now delegates to the real
toast instead of replacing it, and the counting suite clears sonner's
global store before each test. Verified across repeated full-suite runs.

Browser walkthrough

Steps 05-07 still assumed the pre-CL-6342 flow where the sidebar's "+"
minted a workbench on the spot. The harness now opens the picker, chooses
"Just start talking", and presses Create.

Proof

All three walking-skeleton steps pass locally against a scratch Postgres:
bun test scripts/e2e (36 pass, 1 skip, 18 files), bun test test/isolation (41), bun test apps/hub/test (109). bun run typecheck, bun run lint, bun run build, bun run test, the
structural self-tests, and every check:* pass.

Noted, not fixed here

• A run-scoped re-projection lands a new workflow_definition row under
  the agent's own name on every launch, so a tenant's definition list
  grows per launch and "newest sibling by name" resolution can pick a
  run's own derived row. The reuse fix above works around it at the one
  place it broke; the identity model itself wants a ticket.
• CL-6067 reproduced once during live golden-path verification: after a
  hub restart, a reloaded shell stayed disconnected (no Myra sidebar row
  within 15s) on the first attempt and connected on the second. Fresh
  evidence that the reconnect is flaky, not fixed.