Fix:goal continuation cadence fix - part a

[M-Maciej]

Summary

[goal] continuation_delay_seconds (the goal-continuation cadence, added in
#5508) was wired into only one of the two goal-continuation dispatch paths.
The within-turn dispatch hook (goal_continuation_message_if_needed in the
engine) had no wait at all, so a model step that ended with the goal still
active immediately emitted the next continuation prompt. That made the
configured cadence dead on two kinds of sessions:

• CLI-resumed sessions (codewhale resume --last) — they dispatch
  continuation passes through the within-turn hook.
• Host-managed sessions — the engine yields so the host can create the
  next durable claim, so these sessions never reach the cross-turn scheduler;
  every pass goes through the un-waited hook.

Measured (before split) live with a 300 s delay configured: continuation passes fired at
model latency before the fix (~9 s gaps, instant-fire through the within-turn
hook). After the fix, consecutive passes are separated by the configured
quiet window — the observed cadence became ~5 minutes (300 s quiet window
plus the turn itself), versus ~9 s instant-fire before.

This branch applies the wait on the within-turn path, unconditionally:

• goal_continuation_message_if_needed now awaits
  goal_loop::await_continuation_wait before dispatching (biased cancel:
  Esc / steer / host cancel / terminal update_goal all win over a racing
  expiry, and the live goal is re-read only after the wait).
• Host-injected continuation tokens get the same wait in the Op::ContinueGoal
  arm — for host-managed sessions this arm is their only dispatch site.
• The wait stays cancellable, and pause/clear/terminal updates still cancel a
  pending pass.

The flake-hardening test fixes (env-barrier isolation for the route-budget V4
trigger and the prompt prefix-leak tests, SSH-marker neutralization in the
terminal-motion suite, RUSTUP_HOME preservation in the read-only diagnostic
test) are grouped here per the #5535 review guidance (the cadence fix plus the
flake hardening as one small slice).

Rebase note

Rebased onto current main (v0.9.11, merged in #5542). v0.9.11 rewrote
core/engine/turn_loop.rs and the engine event loop; the fix has been
re-verified against the new loop by the regression tests below.

Split provenance

This PR is one slice of a mechanical split: the original combined branch was
rebased onto v0.9.11 and then partitioned along commit boundaries, then each
slice was re-verified independently with the full gate suite. The combined
rebased tree (the union of all slices) also passes the full suite
(13,441 / 0). Partition boundaries worth review attention: the engine
dispatch paths shared with the outbox emit-site work, and the Windows
cfg-gating.

Split map and merge order

The four pieces of the original combined branch are filed as four PRs. All
four are standalone — each is cut from current main and passes its full
gate suite alone:

• (a) this PR — the cadence fix + flake hardening (merge first).
• (b) lifecycle outbox (Feature: local lifecycle event outbox (JSONL + webhook) with turn_stalled / turn_failed events #5531) — same engine dispatch files, different
  lines; builds and passes 13,403 / 0 without this PR. A trivial rebase is
  expected after this merges.
• (c) /relaunch (Feature: /relaunch — switch a running session to the current binary #5532) — the control socket's relaunch verb (d)
  dispatches the /relaunch command through the generic command seam, so
  (d) has no compile dependency on (c): without it the verb returns
  command_error; with it, the verb becomes functional.
• (d) control socket (Feature: the control surface for supervised operation #5533) — zero references to the outbox in either
  direction; no dependency.

One coupling no slice escapes: all four add entries to the same
## [Unreleased] changelog section, so each merge is followed by a small
changelog rebase for the PRs still open. Independently reviewable and
buildable — yes; mergeable with zero follow-up — no.

Testing

<!-- Exact gate commands (including the clippy allow list) are in
CONTRIBUTING.md → "Pre-push verification". -->

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features --locked (warning-free under the CI allow list)
• cargo test --workspace --all-features --locked

Results: clippy clean on stable 1.98.0 under the CONTRIBUTING.md allow
list; full suite 13,375 passed / 0 failed. Focused goal_continuation
suite 13/13 (host-managed sessions wait, non-host-managed sessions wait,
zero-delay dispatch is immediate, cancellation paths).

Checklist

• Updated docs or comments as needed
• Added or updated tests where relevant
• Verified TUI behavior manually if UI changes — the cadence was
  exercised live on running sessions ( yes before split, not after split )
• Harvested/co-authored credit uses a GitHub numeric noreply address —
  every commit is authored under
  130112810+M-Maciej@users.noreply.github.com
• Every commit carries a Signed-off-by (DCO)
• Changelog entry under Unreleased (and the flake-hardening entry
  moved out of the released 0.9.11 section, where an earlier rebase had
  placed it)
• No dead-code-budget change
• The outbox, /relaunch, and control-socket slices from Supervised operation stack: lifecycle outbox, /relaunch, per-session control socket, and the goal-continuation quiet-period fix #5535 follow
  as separate PRs after this one lands

Closes #5534

[github-actions]

Thanks @M-Maciej for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.

[Hmbown]

Reviewed and independently verified locally on your branch:

1. The core design is right. Decision split from dispatch (goal_continuation_allowed), decide-before-wait so a terminal goal never spends the quiet period, biased-cancel wait (Esc/steer/host cancel win over a racing expiry), and — the part I'd call load-bearing — the re-decide on live state after the wait, so a /goal pause, /goal clear, or terminal update_goal landing during the quiet period cancels the pass instead of dispatching a provider request. The Op::ContinueGoal arm correctly gates only host-injected tokens (engine_schedule_id.is_none()); engine-owned tokens keep their scheduler semantics.

2. Ran the focused suite on your branch: goal_continuation 12/12 (host-managed waits, non-host-managed waits, zero-delay immediate, cancellation paths).

3. Ran the repo's single-turn-loop guard (crates/core/tests/single_turn_loop.rs): 2/2 — the turn_loop.rs changes stay inside the one loop.

4. Clippy on codewhale-tui --all-targets on your branch: 0 warnings under the CI allow list.

5. Merge interaction with the 0.9.12 integration branch (#5576): git merge-tree --write-tree between it and this branch exits clean — we also modified turn_loop.rs (compaction decisions, step budget), but in different regions; no conflicts expected either merge order.

The split provenance note and the honest "live-verified before split, not after" checkbox are exactly the review shape we wanted out of #5535 — thank you. Holding the merge for the maintainer's go and for the remaining CI checks to finish.

[Hmbown]

Re-approved at exact head 0903bf6 after the current-main refresh.

Verified locally: all eight M-Maciej commits and their authorship are unchanged; only the expected synchronized Unreleased changelog overlap was resolved. Formatting is clean; workspace/all-targets clippy is clean; goal_continuation passes 12/12; the single-turn-loop guard passes 2/2; the contribution-credit checker passes 11/11 plus the live 8/8 range check. Fresh GitHub CI is green across required Linux, Windows, macOS, lint, and version lanes, with safety, OpenHarmony, DCO, credit/metadata, and advisory checks also green. The PR now closes the exact bug issue #5534.