[kantord]

Summary

Applies the small, independent fixes from #2303 ("Tiny follow-up fixes"), collected from the post-merge reviews of the Studio alignment stack (#2225 to #2256). Each fix that changes behavior has a test, and each new test was checked to fail without its fix.

Part of #2303. It does not close it: the larger bundles and the items that need a design call stay open there (see "Not included").

What changed

#2303	Fix
A, #2225	One token formatter: the session controls use turn-stats' formatTokens, and its false "avoids precision loss" comment is corrected.
A, #2225	The stale "D7-43" and "Phase 7" comments in lib/error-message.ts are gone.
A, #2226	Redundant px-4 / py-3 classes removed from the memory facts table (TableHead and TableCell already default to them).
A, #2232	The Learning page shows a failed background refresh when earlier rows are loaded. Before, a decided row could keep looking pending with no message.
D, #2253	The queue-edit box ignores the Enter that commits an IME composition.
D, #2233	The user docs no longer say the chat header shows "Waiting for authorization". That cue ends when the run parks, and the in-chat status and Review authorization button remain.
E, #2254	A tool row's accessible name includes its status (Open Bash details, failed).
E, #2254	Tool-call previews read only a bounded head of a huge argument or output, and don't parse JSON over 256 KiB.
E, #2255	Forking a chat from the list no longer pulls the reader away from the chat they moved to while the fork ran.
E, #2256	The breakpoint guard also bans min-sm: and min-md:, which compile to the same media queries as sm: and md:.

Behavior changes to know about

• The session controls' token labels now read 1.0k / 1.0M, matching the turn stats, instead of 1k / 1m.
• Tool rows' accessible names change, so a test or script that looks them up by the old exact name needs the status suffix. Two existing tests were updated.
• A tool output over 256 KiB that is JSON is now previewed as text, not as N items or keys: .... Built-in tools are capped at 25 KB, so this only affects very large MCP results.
• A preview of a cut text always ends with an ellipsis, even when the part read was mostly whitespace.

Not included (left open in #2303)

• Sections B and C: the larger bundles (plan-verdict hook, PendingAskCard, schedule detail refresh and focus fixes, and so on). They are refactors or behavior changes, not tiny.
• Optional items that need a design call or more UI: focus after the approval panel closes, gating "Steer agent" on the daemon capability (with its notice), scroll-follow pulling the reader down when they expand something, hidden labels in copied selections, and "just now" for <1m ago.

Verification

Run from apps/:

• pnpm exec biome check .: clean.
• pnpm typecheck: clean.
• pnpm test: 169 web files (1,419 tests) and 26 server files (210 tests) pass. One run had a failure in chat-composer.alignment.test.tsx (a draft assertion after Enter) that did not reproduce in two full reruns or when run on its own. I did not find its cause.
• pnpm generated:check: no drift.
• task docs:check: passes, including the strict link check.

Not run: browser or e2e journeys.

Fully or partially written by an AI agent.

🤖 Generated with Claude Code