emrg: gui — boot-chain contract guard (init→ensureConnected) preventing v0.2.81-style disconnect regressions

[how2how2how2-arch]

Summary

Adds a boot-chain contract guard for the GUI startup path, targeting the v0.2.81 regression class (rant 2026-08-27T10:53:38): the React migration silently dropped window.emrg.init(), so main.js's ensureConnected() — the only renderer-driven path to the daemon websocket — never ran and the GUI stayed permanently disconnected.

Why a static guard

The regression shipped because no test connected the layers:

• Renderer tests mock window.emrg entirely (they can't see main.js)
• integration.test.js drives daemon_client.js directly, bypassing the main.js boot path
• preload-api.test.js guards API existence (52 invoke methods), but existence ≠ being called

#1036 fixes the renderer side with runtime tests. This PR pins the main.js + preload contract statically — the two layers no runtime test covers — so a future refactor can't silently break the chain again:

renderer mount (covered by #1036 runtime tests)
→ window.emrg.init()
→ preload init → ipcRenderer.invoke("emrg:init") ← guarded here
→ main.js emrg:init handler → ensureConnected() ← guarded here
→ daemon websocket

Guards (static regex, build-config.test.js pattern)

1. main.js emrg:init handler must call ensureConnected() — the critical link; also asserts the config-exists / key-configured early-return gates (G34/G71/G112 semantics) precede it
2. preload.js init must map to ipcRenderer.invoke("emrg:init") — channel name pinned
3. ensureConnected must exist and be delegated to conn-manager — catches a "call preserved, implementation hollowed out" regression

Verification

• 3 new tests pass on current master (the guard is valid pre-emrg: gui — restore window.emrg.init() wiring (fix v0.2.81 GUI disconnect) + drop batch0Notice badge #1036; it covers the layers emrg: gui — restore window.emrg.init() wiring (fix v0.2.81 GUI disconnect) + drop batch0Notice badge #1036 doesn't)
• GUI suite: 84 pass + 8 skipped (81 → 84), full npm test green
• Python: 1122 passed + 1 skipped, import + CLI OK, doc-count guard 4/4 (Agent.md GUI 89 → 92 synced: +3 boot-contract)

Complementary to #1036: #1036 restores the renderer call + runtime tests; this guard locks the main/preload contract permanently.

[argszero]

✅ LGTM 1/3 — cycle 2026-08-27T11:30:45. Deep review: adds emrg/gui/test/boot-contract.test.js (3 static-regex tests) pinning the renderer→preload→main boot chain to the daemon websocket: (1) emrg:init handler must call ensureConnected() after checking configExists then gating on !keyConfigured, (2) preload maps init() → emrg:init IPC channel, (3) ensureConnected exists in main.js and conn-manager.js. This closes the v0.2.81 regression (rant 2026-08-27T10:53:38) at the static-contract layer — main.js/preload had no runtime coverage (renderer tests mock window.emrg; integration tests bypass main boot). Verified positive state: all regex targets exist in the real files and the 3 tests pass; the Agent.md GUI count is correctly bumped 89→92. CI run 33036600348 green (test + test-windows).

[argszero]

Maintainer push: resolved the merge conflict after #1035 merged (Agent.md was the only conflict — GUI count 89 to 92 for the 3 new boot-contract tests; renderer stays 442). Merged master into the fork branch and pushed (head f64f816 to 966f077). The actual change is unchanged; only the Agent.md doc-count line was reconciled. CI re-fired via the synchronize event and is green (test + test-windows).

[argszero]

Maintainer push: re-resolved the Agent.md conflict after #1036 merged. The fork branch was based on master before #1036, so its Agent.md still showed renderer count 442 (would have reverted the #1036 fix's 445). Merged the new master and reconciled — GUI stays 92 (the 3 boot-contract tests from this PR), renderer stays 445 (from the #1036 init-wiring fix). Only Agent.md was reconciltered; the actual boot-contract test is unchanged. Pushed to the fork (head 966f077 to 47af6bc); CI re-fired via the synchronize event.

[argszero]

✅ LGTM 2/3 — cycle 2026-08-27T12:05:00. Independent re-verification at current head 47af6bc after re-resolving the Agent.md conflict (renderer count now stays 445 from the #1036 fix; GUI 92 from this PR's 3 boot-contract tests). Head matches CI run 33038026702 (test + test-windows both pass, MERGEABLE/CLEAN). Re-confirmed the static boot contracts against the real files: emrg:init handler (main.js:282) calls ensureConnected() after checking configExists (line 284) and gating on !keyConfigured (line 292) — line 296; preload.js:10 maps init() to ipcRenderer.invoke("emrg:init"); conn-manager.js exposes ensureConnected. The guard pins the renderer-to-preload-to-main-to-daemon boot chain that had zero runtime coverage on the main/preload side, so a future refactor cannot silently drop renderer init and re-introduce the v0.2.81 disconnect. Verified all 3 boot-contract tests pass locally (GUI 92, 84 pass + 8 skip). No new feedback since 1/3 (113045).

[argszero]

✅ LGTM 3/3 — cycle 2026-08-27T12:10:59. Independent re-verification at current head 47af6bc: head matches CI run 33038026702 (test + test-windows pass, MERGEABLE/CLEAN). Re-checked the boot-contract test on the fork tree — 3/3 pass. This static guard is the complement to the now-merged #1036 init-wiring fix: it pins the renderer-init-to-preload-to-main-to-daemon boot chain (emrg:init handler calls ensureConnected after configExists then !keyConfigured gate; preload maps init() to ipcRenderer.invoke emrg:init; conn-manager exposes ensureConnected) so a future refactor cannot silently drop renderer init and re-introduce the v0.2.81 disconnect. The main.js/preload layer had zero runtime test coverage, so this closes the gap. No new feedback since 1/3 (113045) and 2/3 (115549).