Extract shared token estimation helper, migrate all CHARS_PER_TOKEN sites

[avs-io]

Problem

CHARS_PER_TOKEN is defined independently in 7 files (src/context-tree.ts, src/context-budget.ts, and the kiro / cursor / warp / codex / cursor-agent providers), with Math.ceil(len / CHARS_PER_TOKEN) re-derived at each site. Any future refinement of the heuristic would silently apply to some providers and not others, skewing cross-provider comparisons. (#644)

Fix

New leaf module src/token-estimate.ts exporting CHARS_PER_TOKEN and estimateTokensFromChars(chars); all 7 local constants deleted and every call site migrated. One additional hardcoded / 4 site in src/context-tree-codex.ts (same heuristic, just never named) is migrated too.

Behavior-identical by construction: every site already used ratio 4 with Math.ceil, and the two sites with extra semantics keep them — cursor-agent's non-positive guard is preserved, and context-tree-codex's fractional decoded-size input goes through the same ceil(chars / 4).

The module imports nothing from src/, so providers stay lazy-loadable.

Tests

• New tests/token-estimate.test.ts pins the heuristic: ratio value, ceil rounding, zero/boundary inputs. The ratio can now only change in one place, guarded by one test.
• npx tsc --noEmit clean.
• Full suite: npx vitest run — 130 files passed, 1713 tests passed, 0 failed.

Closes #644

[iamtoruk]

Verified in depth: no stray CHARS_PER_TOKEN or hardcoded /4 sites remain, the ratio pin tests fail when the constant changes, the cursor-agent non-positive guard is preserved, and a 30-day all-provider report on real data is byte-identical between main and this branch, so the behavior-identical claim holds empirically. The two local test failures during verification are pre-existing on main (app/electron/cli.test.ts expects a built dist/cli.js) and unrelated.