fix: fill context bar against soft limit, reconcile token figure

[tmck-code]

Context

The context fill bar stopped filling to 100% at the soft limit. Commit f514d3a ("use used_percentage for context bar fill") switched the bar to fill from the host-supplied used_percentage, which Claude Code measures against the full context window (e.g. 200K) rather than the soft limit (150K, the compaction-risk threshold). As a result the bar only reached 100% when the real window filled, and the soft_limit argument became dead code inside _ctx_fill_ratio.

A second, related inconsistency: the bar's displayed token figure used total_input_tokens while the fill came from used_percentage, so the number and the fill could disagree (e.g. 428 shown next to an 11%-filled bar).

Changes

• Restore soft-limit-relative fill while keeping f514d3a's improvement (host-accurate, input-only token count). _ctx_fill_ratio now converts used_percentage back to an absolute count via context_window_size, then divides by soft_limit. Reaches 100% at the soft limit again; the window-headroom figure stays as the parenthetical (N%).
• Reconcile the displayed figure with the fill. New _ctx_used_tokens(ctx) helper is the single source of truth — the displayed token figure, fill length, risk-zone colour, and (N%) all derive from it, so they can never drift. Prefers host used_percentage rescaled to the window; falls back to total_input_tokens; clamps to >= 0; excludes output tokens.
• Docs: updated CONTEXT.md — the bar's length now tracks the soft limit (correlated with colour), window capacity shows only as (N%); added a Used Tokens glossary entry and reworked the example dialogue.
• Snapshots: regenerated the three claude-dark golden fixtures (fill shifted 8%->11% on the example; wide figure now 16.0K).

Checklist

• Tests added or updated for new/changed behaviour
• (bug fixes) a regression test now covers the situation so it can't come back

Screenshots / recording

Before	After

Benchmark

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	65.6 ± 1.6	62.2	70.3	1.00
PR	68.6 ± 4.9	64.2	99.3	1.05 ± 0.08

Within noise — the PR adds only one small helper call; the 1.05x reflects outliers (PR σ 4.9ms, max 99ms).

System info

Key	Value
OS	Linux archlinux 7.0.10-arch1-1 #1 SMP PREEMPT_DYNAMIC Sat, 23 May 2026 14:21:20 +0000 x86_64 GNU/Linux
Claude Code	2.1.165 (Claude Code)
Terminal	TERM=tmux-256color TERM_PROGRAM=tmux SHELL=/bin/bash COLORTERM=truecolor
Locale	LANG=en_AU.UTF-8 LC_ALL=
Python	Python 3.13.1
uv	uv 0.11.14 (x86_64-unknown-linux-gnu)

🤖 Generated with Claude Code