Keep prompt history byte-stable so the cache prefix survives

[jurgenwerk]

Replaying two real assistant sessions turn by turn through the prompt builder and byte-diffing consecutive prompts showed 40-60% of session cost going to avoidable prompt-cache misses. The prompt's cache design (volatile context trailing a moving history breakpoint) is sound, but three things kept breaking the prefix:

tools ──► system+skills ──► history ──► [cache marker] ──► volatile context
  │             │               │
  │             │               └── attachment/tool-result content was dropped
  │             │                   retroactively when a newer version arrived
  │             │                   → history bytes changed mid-prompt
  │             └── survived (its breakpoint was the only hit on such turns)
  └── changed when a card attach inserted a per-card tool → full re-bill

The main change: a message's attachments now render from that message's own snapshot alone. Previously an attached card or file lost its content in the prompt once a newer version was attached later, and read-file tool results lost theirs the same way — rewriting already-sent history and re-billing everything after the change point at full input price on every subsequent turn. Superseded content now stays, and the attachment headers tell the model that a later attachment of the same card/file supersedes earlier ones. The carried content is billed at the cached-read rate, an order of magnitude below the re-bills it replaces. Replaying the same sessions with this change shows every turn diverging only at the trailing context message, as designed.

Two supporting ai-bot changes: Anthropic-model requests carry an OpenRouter provider preference for Anthropic (caches live per provider, and one observed request was routed to a cold instance mid-session), and per-turn usage now records the serving provider and generation id so a cache miss is attributable from the room timeline.

Known, accepted misses not addressed here: the first turn of a room runs before skill tools are discovered (small prompt, one-time), and attaching a card inserts its patch tool (semantic, once per session). Making per-card patch schemas byte-stable would need a tool-definition redesign and is left out deliberately.

[github-actions]

Host Test Results

    1 files      1 suites   2h 3m 42s ⏱️
4 192 tests 4 178 ✅ 14 💤 0 ❌
4 211 runs  4 197 ✅ 14 💤 0 ❌

Results for commit 0ae799e.

Realm Server Test Results

    1 files      1 suites   15m 47s ⏱️
2 173 tests 2 173 ✅ 0 💤 0 ❌
2 253 runs  2 253 ✅ 0 💤 0 ❌

Results for commit 0ae799e.