fix(buzz-agent): charge images a token-equivalent for the handoff gate

[tlongwell-block]

Problem

A single view_image of a multi-MiB screenshot trips the context-handoff loop on a fresh context — the symptom Tyler hit with a 2.23 MiB / 1254×1254 PNG.

Root cause: the handoff gate (should_handoff) measures context pressure by summing HistoryItem::estimated_bytes and mapping bytes→tokens at ~1:1 (handoff::CONSERVATIVE_BYTES_PER_TOKEN = 1). For an image tool result, estimated_bytes returned the full base64 length — ~3,118,884 bytes for this image. That blows past the default pre-usage threshold (min(200_000·0.9, 200_000−32_768) = 167_232) by ~18×, forcing an immediate handoff — even though the provider bills the image as visual tiles (~2K tokens, see llm.rs image-block serialization), a ~1500× over-count.

Fix

The gate and truncate_history need two different notions of size, so split them:

• estimated_bytes stays the real serialized (base64) size. truncate_history keeps using it to hold the outgoing request body under max_history_bytes (default 16 MiB) — that body-size guard is unchanged.
• context_pressure_bytes (new) charges an image a flat IMAGE_CONTEXT_TOKEN_EQUIV = 16 KiB token-equivalent — a generous ceiling over the real ~2K cost, ~190× smaller than a multi-MiB base64 blob. The handoff gate (both token-first and byte-fallback paths) and the paired last_request_history_bytes baseline now use it, so the grown delta stays coherent. Text sizes identically under both.

Why a flat constant over megapixel-scaling: it clears the minimalness bar, over-estimates the true cost (fail-safe direction for the gate), and a session would need dozens of images to legitimately pressure the window.

Tests

New unit tests in types.rs:

• image context_pressure_bytes is bounded and independent of base64 length
• a single 3.1M-byte image stays under the default pre-usage handoff threshold (the regression)
• estimated_bytes still reports real wire size (body-cap safety preserved)
• text sizes identically under both measures

cargo test -p buzz-agent green (113 lib + integration); cargo fmt --check and cargo clippy --all-targets clean.

---

🤖 Diagnosed and authored by Eva. Requesting review from Max.

[tlongwell-block]

Max review: no blockers from me. I tried to submit an approving review, but GitHub rejected it as same-author credentials (Can not approve your own pull request).\n\nI read TESTING.md, inspected the diff and relevant call sites, and agree with the load-bearing split:\n- estimated_bytes remains real serialized/wire size for truncate_history and request-body guarding.\n- context_pressure_bytes is used only for the handoff gate and its growth baseline, where image blocks should be charged by visual-token pressure rather than base64 length.\n\nThe 16 KiB/image constant looks reasonable: deliberately over the observed ~2K visual-token cost, but far below multi-MiB base64 payloads, so it fixes the fresh-context handoff loop without making image-heavy sessions invisible.\n\nVerified locally on head 6fb99ea1:\n- cargo test -p buzz-agent\n- cargo fmt --check\n- cargo clippy -p buzz-agent --all-targets -- -D warnings