emrg: release v0.3.18 — six commits since v0.3.17, two of them GUI defects the host is still running #2358

Origin: host instruction — 2026-10-15T22:15:21+08:00, session s_260727_1103_866d: "not 3.14 — I mean: after all the rants are done, publish another new version" (translated from the Chinese original; resolvable on this host with scripts/find-host-message.py --pattern '发布新版本', exit 0). The same instruction is stated at 2026-10-15T22:13:09.

That condition holds now: every emrg rant is completed — the queue reads 0 pending, 0 in_progress across all 66 entries — and the two rants filed on 2026-10-17 as part of that set had their fixes land on master after the last tag. This issue is the release those fixes are waiting for.

Baseline, measured (2026-10-18T06:57 +08:00)

• Latest release: v0.3.17, published 2026-10-17T11:43:38Z — check-release-published.py v0.3.17 reads OK: build run 39029795077 green, draft=False prerelease=False, 11 assets, github Latest: v0.3.17.
• The tag sits on 0b8a2e77; master is 6 commits ahead at e268b4c7.
• The host's install reads ~/.emrg/install/version.txt = 0.3.17, so none of the six is in force on this machine yet.

Why this release matters beyond the count

Two of the six close defects the host sees in the GUI and has been running since before v0.3.17:

• e268b4c7 — a tool-call group's height flipped at every tool boundary, because the expansion state was derived from "is any row still running". Five calls meant five jumps, and the scroll position went with them. The state now has one writer, the host's click.
• 1e9b7e77 — a reopened session could not render what the live transcript had shown: list_history emitted only user/assistant messages and tool results, so system records and compaction summaries reached no client, and a cleared session replayed empty.

db3cf8b2 matters for a different reason: a scheduled task's extra_prompt was never read at all, so every hand-written instruction in tasks.yml was silently inert. It is on master and unreleased, which means every scheduled cycle running on the installed 0.3.17 is still not reading it.

Stages

1. Bump — scripts/bump-version.py 0.3.18 rewrites every version declaration at once; tests/test_version_sync.py proves they agree. PR it, take three approvals from different cycles, squash merge. A release PR is never self-merged.
2. Tag — git tag -a v0.3.18 -m "emrg v0.3.18" && git push origin v0.3.18. The tag is the only trigger of build-release.yml.
3. Verify — scripts/check-release-published.py v0.3.18: green across the platform matrix, published as Latest (not draft, not prerelease) with the full asset set.

What this release does not do

• It changes no behaviour by itself: a version bump plus a tag. Every behaviour change in it is one of the six commits above.
• It closes no issue other than this one. The two open issues (emrg: input history belongs to the session in the daemon, not to a client process #2349 input history, emrg: an issue must stand on its own, and the feedback-to-issue pairing must be read from where it is durable #2352 the issue/feedback pairing) have no merged work and are untouched by it.