[jvmenen]

Problem

The dashboard request log showed SSD_PREFIX_MISS in orange for almost every miss. The cold tier reports that reason whenever it holds nothing reusable, and the session bank copied it over its own reason. So a new conversation, a prompt that shares less than the reuse minimum (512 tokens) with what is stored, and a turn that was evicted from memory all looked like an SSD fault, although the SSD is only the last place searched and nothing failed.

Change

Server: a plain ssd_prefix_miss is now reported as what it meant:

• new_session: the bank holds nothing.
• prefix_divergence_at_token: other conversations are stored, but the shared prefix is below the reuse minimum.
• evicted: this conversation's own earlier turn left memory and is not on the SSD.

Dashboard labels, with the precise reason in the tooltip:

reason	label	colour
new_session	NEW CHAT	grey
prefix_divergence_at_token	NO SHARED PREFIX	grey
evicted	EVICTED	yellow
ssd_restore_error:*, ssd_missing_*, ssd_prefix_no_recurrent_boundary	SSD ERROR	orange
ssd_prefix_miss (older servers)	NO SSD MATCH	grey

Compatibility

No new reason values; all three already exist in CacheMissReason. ssd_prefix_miss is still reported when the memory lane refused a prefix for another reason (the cause is in last_ram_miss_reason), and real SSD failures keep their reasons. The dashboard still reads ssd_prefix_miss from older servers.

Tests

• Server: a reason per case in tests/test_ram_miss_reason.py (new conversation, no shared prefix, own turn evicted, another conversation not treated as the owner, SSD failure passed through); three existing assertions updated to the new reasons. Ran with test_cache_bank, test_cold_tier_disk_usage_scan, test_prefill_reread and test_session_bank: all pass.
• Dashboard: bun test (label and colour per reason, rendered row with tooltip), tsc --noEmit and bun run build clean. The built assets are in a separate commit.