[SYS-002] Agent --resume ignores MEMORY.md: complete context loss on session resume, decision instability, destroyed live test

Summary

Agent resumed into an active session (river permit sniper automation) and completely ignored its own MEMORY.md and all project CLAUDE.md instructions. The result was 2+ hours of wasted time during a time-critical live test window, reinventing already-solved problems, and a lost real-world opportunity (Desolation Canyon river permit drop).

Environment

• Claude Code v2.1.76 (latest as of Mar 15, 2026)
• Node.js mode (switched from Bun Feb 26 — zero crashes since)
• Windows 11
• Agent mode with @river-permit-sniper named agent

What Happened

1. Agent was resumed into an existing session
2. Agent had no knowledge of prior work — acted as if starting fresh
3. Agent flip-flopped on a trivial fact (which Chrome profile to use) 4+ times in the same session: RecreationBot → Account1 → TestAccount → ChromeProfile_TestAccount → Account1 → TestAccount
4. Each time the user corrected the agent, it over-corrected past the right answer
5. Agent spawned sub-agents to "update memory/docs" — those updates apparently didn't persist or weren't loaded back
6. Agent rebuilt login logic that already existed and was tested, tried failure modes already proven not to work
7. Session ended without completing the task

Impact

CRITICAL. This was a live permit drop test — one shot per day at 7:00 AM PT. The agent's dysfunction consumed the entire window. The permit test slot and the actual permit drop opportunity were both lost.

Failure Mode Analysis

The core issue: MEMORY.md was not loaded on --resume. The agent's memory system is only useful if the contents are injected at the start of each session. Without it:

• The agent has no stable reference for prior decisions
• User corrections get overridden by the agent's priors
• Already-solved problems get re-researched and re-broken
• Time-critical automation becomes unreliable

Relation to Known Issues

• #9796 — CLAUDE.md ignored post-compaction: Same failure mode (instructions not loaded), different trigger (compaction vs. resume)
• #7515, #5528, #4554 and related sub-agent instruction inheritance cluster: Agent memory not surviving context boundaries
• v2.1.76 PostCompact hook (addresses #14258/#17237): Supposed to help with instruction loss — did not prevent this failure

Expected Behavior

On --resume, the agent should:

1. Load and act on MEMORY.md before processing any user turn
2. Treat MEMORY.md contents as ground truth for prior decisions
3. Verify solutions exist in the codebase before proposing new ones

Requested Fix

1. Confirm whether --resume force-injects the agent's MEMORY.md
2. If not: add MEMORY.md injection to the --resume startup sequence
3. If yes: investigate why the model ignores it despite injection
4. Consider a "did I already build this?" pre-check in agent mode before any file creation or modification