Count reasoning, images and fixed overhead in the context estimate (dirge-qobx.1)

[yogthos]

The fold trigger reads the provider's prompt count. The fold's own accounting read chars / 4 over messages, counting only text blocks and tool-call arguments. Three things that ship with every request counted zero:

• replayed reasoning — a thinking block is echoed back inside the assistant turn for every provider except OpenAI (provider_rejects_reasoning_echo), and nothing ever strips a stale one. On a long reasoning-heavy run it is the largest single term in the prompt.
• images — the transcript holds an asset id, the request carries ~1.5k tokens of billed area.
• the system prompt and every tool schema — ~16k tokens for the built-in surface, ~33k with MCP servers (measured in compact_schema), and not in messages at all.

So a request the provider charged 82% of the window for read as 25% at the turn-start tier: the 0.90 fold never fired on anything, the 0.60 result-cap tier tightened late, and a fold could report context compacted: 63800 → 63479 against a 204,320-token request — which is what a live run did, one turn before stopping mid-task.

• block_chars prices thinking text and image blocks (IMAGE_TOKENS_ESTIMATE, a flat 1.5k — the transcript records no dimensions).
• The fixed overhead is re-derived after each response as prompt_total − estimate(messages we sent) and added back at the two pre-send tiers. Last observed, not a running maximum: the tool surface changes with /model, an MCP server connecting, or a prompt-layer swap.

Also dirge-qobx.6, because the derivation above depends on it: on Anthropic input_tokens is the uncached remainder, not the prompt (TokenUsage's own doc says so). With a warm prompt cache every tier under-read the prompt by the entire cached prefix — no fold at 0.75, none at 0.78, no exit-with-summary at 0.80, and the first sign of trouble was the provider refusing the request. TokenUsage::prompt_total normalizes the two conventions, every decision site reads it, and Session::record_token_usage follows the same convention so the gauge keeps describing the fold it warns about.

Follow-ups in the epic (dirge-qobx): strip stale reasoning at the provider boundary (qobx.2), stop ending the run on a prune-only fold (qobx.3), fold when an autonomous stretch has no user turn (qobx.4), log the silent skip (qobx.5).