fix(cli): stop unchanged progress feedback loops [risk:high]

[zxch3n]

Risk: 🔴 high | Confidence: high — changes CRDT history write behavior; real-Mirror regression tests and independent review cover convergence and subsequent progress updates.

Related issue

Same-repository fix; no separate tracking issue.

Problem / pressure

Nested create operations (A → B → C) can keep the CLI coordinator reconciling indefinitely. Updating B's progress notifies A's target subscription even when progress has not changed: returning the original history inside Mirror.setState still emits notifications.

Summary

Skip unchanged or absent progress before entering updateHistory. Recompute actual writes against the latest history. Preserve duplicate-ID recovery and monotonic status merging. Add real-Mirror coverage and keep the notification invariant in AGENTS; move explanatory connection-lifetime material into the adjacent README to respect the document size limit.

Before / after

Before	After
Nested progress repeatedly wakes its own coordinator; the regression reaches its 20-scan safety guard.	The coordinator quiesces after two scans.
Returning unchanged history still notifies Mirror subscribers.	Unchanged progress never enters Mirror; later child completion still updates and settles its card.

Test plan

• Full standalone OSS pnpm check passed on the implementation commit: workspace typechecks, lint, tests, i18n and boundary checks.
• 145 orchestration/progress tests passed against base bad3c75; the old implementation failed both new regression cases.
• Independent race/correctness, security/scope, and test reviews found no P0/P1; correctness was re-reviewed after refreshing the base.
• CI-equivalent static steps pass locally: 45 contribution-policy tests, pnpm run docs check --base origin/main, pnpm format:check, pnpm typecheck, and pnpm check:quick. Document maintenance and formatting are separate from pnpm check.
• A separate 0.92.0-based diagnostic package passed the feedback reproduction using its embedded Electron runtime. Affected-user workspace validation remains pending.

Context handoff

Instructions for reviewing agents

• Review focus: Check operation-progress-history.ts and the real-Mirror nested coordinator regression for stable-state quiescence and subsequent completion updates.
• Decisions to challenge: The preflight read suppresses no-op writes, while the actual updater recomputes against current history; ensure duplicate-ID repair remains recoverable.
• Plausible failures / evidence gaps: Look for missed concurrent updates or feedback paths; the affected user's original workspace has not yet confirmed recovery.

Authoring context

• User goal / directives: Fix a reproduced CLI progress feedback loop, add a regression test, and submit the change using the repository PR template.
• Constraints / non-goals: Preserve operation semantics, persisted schemas, real completion updates, and cloud/local boundaries; exclude captured user data and installation artifacts.
• Risk-bearing decisions: Suppress a CRDT write before Mirror notification; preserve the latest-state updater and existing duplicate repair rather than globally changing Mirror behavior.
• Destructive or irreversible behavior: No migration, cleanup, deletion, or deployment is introduced; existing history repair and recovery behavior remain intact.
• Deliberately not done or tested: No broad scheduler redesign or production rollout; validation on the original affected workspace is pending.
• Unknowns / confidence: High confidence in the reproduced loop fix from real-Mirror tests and independent reviews; no claim that every reported CLI stall has the same cause.