Render a semantic activity state in the status ticker, not raw tool names

[TheGreatAxios]

Summary

• CL-5732: the bottom-left status ticker rendered whatever raw tool identifier was currently executing (run_shell, grep, MCP names, ...), leaking plumbing vocabulary into a product surface and duplicating what the transcript already shows.
• resolveTurnLabel in src/tui-opentui/session-chrome.ts now returns a member of a small closed set — thinking, planning, researching, building, working, waiting, stalled, stopping — via a single tool-to-state table (TOOL_ACTIVITY_STATES) with an explicit working fallback for anything unmapped (built-in, MCP, or plugin). Consumed at the actual leak site inside paintPhaseAt in runtime-bridge.ts.
• Stalled reuses the existing isStalledForDisplay signal from stall-watchdog.ts — no second notion of stuck. Blocked/waiting still derives from the existing gate signal (blockedGateCount / status === "blocked"); only the rendered literal changed from blocked to waiting so it reads distinctly from working.
• The closed set is now compiler-enforced, not just test-enforced: AppShell.lockupPhase and LockupFrame.phase in shell.ts are typed ActivityState | null, so a raw identifier reaching setLockupFrame is a type error. lockup.ts's LockupInput.phase deliberately stays a generic string — its own tests feed it arbitrary CJK/astral strings to check column-width math, unrelated to the activity vocabulary, and the leak boundary is already closed one layer up at the setLockupFrame call site.

Scope decisions (recorded per review)

• Rendering case: lowercase, unchanged. The operator's ask was for human words over raw identifiers; the ticker's existing house style (thinking, working, etc.) is lowercase and unpunctuated by design (see resolveTurnLabel's doc comment), and a single capitalized word would read as a mistake against the rest of the chrome. Kept lowercase.
• agent-progress.ts's ${elapsed} · ${tool} is intentionally out of scope. That string feeds the sub-agent panel rows, not the prompt-box lockup slot this ticket targets. Per the operator: raw tool names on sub-agent rows are desired — that panel is where an operator watches what a delegated agent is actually doing (grep, etc.), and collapsing it to a generic activity word would remove information they want there. The lockup slot answers "is it alive"; the sub-agent rows answer "what is it doing." Left unchanged. No follow-up ticket needed per the operator's ruling.
• src/tui/tool-formatter.ts already holds two tables keyed on the same tool-name space (TOOL_DISPLAY_NAMES for transcript display, plus the MCP humanizer). TOOL_ACTIVITY_STATES in session-chrome.ts is a third, deliberately different taxonomy (activity category vs. display name) over the same key space — they'll want reconciling when the old TUI goes away, not now.
• streaming N tok was deliberately removed from the label, not collateral damage — the closed-set guarantee requires the rendered string to always be a fixed member of ACTIVITY_STATES, which a live token count can't be. The ramp's own cell animation still carries "is it alive" during text streaming (state renders as working).

Test plan

• bun run test — 4212 pass, 1 pre-existing unrelated failure (src/agent/lsp-availability.test.ts, environment-dependent language-server detection, untouched by this change)
• bun run typecheck — clean on all touched files; pre-existing vendor @intx/types version-drift errors remain in src/subagent/run.ts / src/tui/runner.ts, neither touched here
• bun run build
• CI's transcript-long-log-scroll.test.ts failure investigated: reproduced clean-main (19d9710) and this branch locally with the exact CI invocation (bun test ./src ./tests ./evals --randomize --seed 424242) plus isolated repeats of that single file on both. It passed every time on both revisions; the diff does not touch long-log.ts, shell.ts's scroll path, or harness.ts at all. Timing-sensitive/flaky, not caused by this change.

[linear-code]

CL-5732