fix(agent): inject memory + reflections per-turn, not in frozen snapshot (#41)

[mattmezza]

Closes #41.

Problem

In session mode the static system prompt is snapshotted once per session so the cacheable prefix stays stable and the provider's prompt cache hits. But memories and task reflections were baked into that snapshot, so anything extracted mid-session was invisible to the model until /new rebuilt the snapshot. It bites hardest:

1. After compaction — a fact extracted early, then compacted out of raw context and absent from the frozen snapshot, becomes genuinely invisible.
2. Cross-chat — a fact taught in chat B lands in the global store but chat A's open session can't see it until A resets.
3. Reflections — a lesson from a tool failure is most useful on the next similar task, often in the same session.

Fix (the cheap one)

Move memory + reflections out of the frozen snapshot and into the existing per-turn, never-cached preamble (_turn_preamble), which already injects live date/time + the execution plan onto the current user message.

• Static snapshot keeps only truly static content (persona / skills / instructions) — still snapshotted once, cache intact.
• Memory is fetched fresh each turn via memory.format_for_prompt(query=message) — already query-aware — so it is both fresh and per-turn relevance-ranked.
• The preamble rides on the new user turn, which is uncached anyway, so there is no extra cache miss on the static prefix or prior messages. Cost is only the (bounded, top-k) memory block's own tokens, paid on a turn you'd pay for regardless — far cheaper than rebuilding the whole prompt.
• Resolved the issue's open question as preamble for both modes → one code path; injection mode also gets query-relevant-per-turn memory.

Reflections are injected fresh too (still query-less / most-recent-N — making them query-aware is the noted follow-up). The recall_memory tool is the issue's explicitly-deferred phase 2 and is not in this PR.

Self-check

test_mid_session_memory_visible_next_turn_without_new asserts a long-term memory written mid-session appears in the next turn's preamble — without /new — while the static snapshot stays frozen.

Docs

Updated architecture.mdx and memory.mdx to describe per-turn injection (dropped the stale "snapshot holds memories" / "first message is the query" wording).