[omartelo]

An Opus 5 session read 6% of its context as 30% in the footer.

What was wrong

The window the percent is measured against came from a table listing the 1M
models. opus-5 was not in it, so the read fell through to inferring a window
from the token count — the smallest standard size the session still fit inside,
which is 200k until a session passes 200k tokens. Hence 5×.

Adding the missing row would have fixed this launch and left the next one
broken. The table's default was the actual bug: it enumerated the large
windows and assumed small for anything unlisted, so every model release needed a
release here to read right.

What changed

Inverted. Haiku is the 200k exception, everything else is 1M:

func windowForModel(model string, tokens int) int {
	if strings.Contains(model, "haiku") && tokens <= smallWindow {
		return smallWindow
	}
	return defaultWindow
}

The bet is that an unknown id is a model newer than the binary rather than a
legacy one, and it holds — every Claude generation since 4.6 ships at 1M. A
token count past 200k disproves the Haiku guess (a window cannot hold more than
its size) and widens, so a future Haiku at 1M corrects itself.

That collapsed modelWindows, contextWindows and contextWindowFor into the
four lines above: 42 insertions, 70 deletions. The deprecated ids the old
table carried went with them — nothing in a provider's model picker offers them
today.

Two tests changed because the contract did, not to stay green: an unlisted model
now takes 1M instead of the smallest window that fits. The deprecated-model rows
were dropped for the same reason as the code.

The CLAUDE.md ceiling for this went too. That section carries what breaks if
nobody knows it; an approximate readout does not qualify, and the mechanism now
lives in the comment on windowForModel, where someone changing it will read it.

Tradeoff

A session on a still-live 200k model outside the exception (Sonnet 4.5, Opus
4.5) now reads low. The widen guard only ever widens, never narrows. Accepted:
those are not what the providers spawn today, and under-reading a model nobody
runs beats a table that needs a release per model launch.

Not done here, deliberately

Reading the exact window per session. Checked and ruled out:

Source	Why not
Transcript JSONL	schema dump confirms no window field anywhere — only token counts
claude CLI	no models subcommand
statusLine (context_window.context_window_size)	exact, but the slot belongs to the user's own status line
Models API (max_input_tokens)	authoritative, but needs auth + network, and still cannot see a window capped below the model's native size

So a 1M-capable model held at 200k still reads low. Same ceiling as before this
PR, now stated in the code rather than the prompt.

Test plan

• go test ./... — 400 pass
• go vet ./... and gofmt -l . clean
• cross-compile loop: GOOS=linux|darwin|windows go build ./... && go vet ./...
• smoke in task dev: footer reads ~6% on an Opus 5 session, not ~30% — verified manually before merge