fix: use context_window for [1m] model display instead of total_tokens

[KorenKrita]

Problem

Sessions were incorrectly displaying the [1m] suffix on models that only have a 200k context window.

Root cause: sessions.rs was using session.total_tokens() > 200_000 to detect 1M context, but total_tokens() accumulates all token usage across the entire session lifetime, including cache_read tokens. Since cache_read is reused on every turn, a long-running session on a standard 200k model can easily accumulate 200k+ total tokens and trigger a false [1m] display.

Additionally, fmt_tokens always rendered one decimal place, producing awkward labels like 200.0k instead of 200k.

Fixes

1. Use context_window for exact [1m] display (sessions.rs)

// Before (wrong): total_tokens includes cache_read and grows over session lifetime
let is_1m = session.total_tokens() > 200_000 || session.model.contains("[1m]");

// After: only render the literal [1m] suffix for actual 1M windows or explicit model suffixes
let is_1m = session.context_window >= 1_000_000 || session.model.contains("[1m]");

The collector already computes context_window. The UI should not infer a 1M label from cumulative token totals, and it should not treat every non-200k window as a literal 1M window. Codex can report windows such as 258.4k, which are larger than 200k but are not 1M.

2. Drop unnecessary decimal in fmt_tokens (mod.rs)

Before: 200.0k, 1.0M
After:  200k, 1M

Fractional values like 123.4k are preserved. Whole numbers get a clean display.

3. Regression coverage

Added a sessions-panel regression test for a Codex session with context_window: 258_400, verifying that the rendered model label does not include [1m].

Affected files

• src/ui/sessions.rs — use exact 1M detection for the [1m] model suffix and add regression coverage
• src/ui/mod.rs — drop unnecessary decimals in fmt_tokens()

Testing

• cargo test — 137 passed
• cargo clippy -- -D warnings — passed