fix: honor ambient hot-reload and platform-aware badge tests

[1jehuang]

Summary

Fixes #1269. Fixes #1275.

Two independently reviewable, narrow fixes from issue triage:

• [ambient] Enabling ambient in config.toml has no effect until restart, while ambient:status reports enabled: true #1269: Re-read ambient.enabled at the loop gate instead of retaining the startup value. Preserve explicit AmbientStatus::Disabled, startup-only reply pollers, and the existing scheduler/wake policy. Add a disk-config regression covering false → true → false and the explicit-stop override.
• macOS: copy-badge margin test assumes the non-macOS Alt-label width #1275: Construct expected shortcut text with the configured/platform modifier label and Unicode display width. Keep unaffected-row and exact-fit assertions, and fix the neighboring fixture's literal Alt. No production badge rendering changes.

Commits: 2200140 (#1269), 180e95d (#1275).

Verification

Performed on Linux x86_64 through the coordinated self-dev test/build queue:

• Reproduced the original badge assertion with an isolated JCODE_HOME and [display] copy_badge_alt_label = "⌥": 18 != 16 before the fixture fix.
• scripts/dev_cargo.sh test --offline --locked -p jcode-tui --lib copy_badge -- --test-threads=1: 16 passed with default Alt, and 16 passed with configured ⌥, including platform-default and trimmed-override tests.
• Ran each of the two changed badge fixtures separately with --exact --test-threads=1 under each label: 4 exact invocations passed on the final formatted tree.
• scripts/dev_cargo.sh test --offline --locked -p jcode-app-core --lib ambient::runner:: -- --test-threads=1: 6 passed, including the new toggle regression and real running-loop hot-reload regression and existing disabled-ambient scheduled-task coverage.
• scripts/dev_cargo.sh build --profile selfdev -p jcode --bin jcode: passed via selfdev build.
• git diff --check and rustfmt --check --edition 2024 on all three changed files: passed.

Measured behavior follow-up

Additional regression coverage in c515991 exercises the actual run_loop, not just the extracted gate. It starts disabled, edits config.toml to enable ambient without invalidating the cache, and observes the loop enter Paused for an active user session after nudging it. This exercises file fingerprint reload and the production gate without model inference.

Restoring the old startup-snapshot gate made this exact test fail after its 3-second observation deadline. Restoring the fix made it pass, along with all six ambient runner tests. This establishes a behavioral improvement at the real loop boundary, while not claiming a full CLI/daemon or native macOS acceptance run.

Limits and non-goals

• Native macOS execution is still needed. Linux ⌥ coverage exercises the reported width difference, not the macOS OS/runtime.
• Ambient edits take effect at the next loop wake. This does not promise a 500ms wake or interrupt an already-running cycle. Reply-poller and scheduler hot-reconfiguration remain out of scope.
• The shared daemon was not reloaded. Runtime assertions above are from the compiled test targets, not an end-to-end live-daemon hot-reload run.
• No full-workspace green claim: independently reproduced the pre-existing Conifer completeness failure in Conifer static fallback lacks context limits for 25 models and fails catalog completeness test #1274 (the same 25 missing IDs), which requires a separate metadata/alias policy decision.
• Existing contributor fixes for swarm list omits model identity for a busy worker #1266/swarm await_members redundantly fetches full history for reported members #1264/Daemon startup fails when the runtime directory does not exist #1252/Shift+Enter requires extended-keys always because Jcode does not opt in under tmux #1250 remain in their own PRs. No duplicate implementation or unrelated branch is included.

Hosted CI follow-up

This PR is draft pending hosted checks and native macOS validation. The CI run, Windows smoke, and iOS workflow all failed before creating any jobs. gh run view reports a likely workflow-file issue, and these unchanged base workflows each contain duplicate top-level env maps. The existing repair is tracked by #1191 / PR #1193 and is not duplicated here. Require Linked Issue passed. Local test/build results above remain valid, but this is not green hosted CI.

---

Jcode agent (automated triage), on behalf of @1jehuang