refactor(tui): split 7.8k-line app.rs into cohesive app/ modules

[nhogenson]

Summary

tui/src/app.rs was a 7,807-line file containing the full App state struct, a 5,557-line impl block, the keyboard dispatch tree, and a 950-line test suite. This PR restructures it into app/ with one module per responsibility. Pure structural move — zero behavior change. Every function body is byte-identical to the original; no logic was touched, no warnings fixed, no style "improved" while moving.

Module layout

Module	Responsibility	Lines
mod.rs	App struct, constructor, re-exports, module wiring	~940
types.rs	Pure data model (messages, tool blocks, dialogs, state)	245
keys.rs	Keyboard normalization + key-event dispatch	2,727
mouse.rs	Pointer input, selection, context menu, paste viewer	660
providers.rs	Provider/model/config + import-config flow	442
commands.rs	Slash-command catalog + intercept dispatch	392
turns.rs	Turn lifecycle (snapshots, streaming flush, labels)	176
messages.rs	Transcript ops (add/replace/push, notifications, scroll)	162
prompt.rs	Prompt input state machine + voice hold-to-talk	158
views.rs	Overlay/dialog orchestration + MCP view state	405
run.rs	Main event loop, query events, jump-to-error	628
tests.rs	Test suite (moved whole, use super::*)	956

Public API unchanged — crate::app::{App, ...} re-exported from mod.rs exactly as before; zero ripples into render.rs, messages/, or lib.rs.

Visibility note (the one deliberate adjustment)

Private methods moved across module boundaries became pub(super) — restoring exactly the visibility the single-module layout provided. Of the 79 items lifted: 48 have genuine cross-module callers; 31 are used only within their defining file and are candidates to return to private during the phase-two interface work below.

Placement decisions

• try_copy_to_clipboard stays in mod.rs — it's crate-level public API, re-exported by lib.rs (pub use app::try_copy_to_clipboard).
• Animation/mode helpers (tick_rustle_pose, cycle_agent_mode, apply_theme, …) stay near the constructor for now; their callers live inside app, and regrouping them belongs with the interface work.
• MCP plumbing in views.rs mirrors the original file's conflation; splitting it out is flagged for phase two.

Known follow-up (phase two, separate PRs)

Honest caveats this PR does not hide:

• keys.rs remains the largest file — handle_key_event alone is ~2,300 lines and is the next target.
• Submodule impl blocks still share the full App state; phase two replaces &mut self-everywhere with narrow interfaces and encapsulated state, and privatizes the ~31 lazy lifts.
• cargo fmt stays advisory per repo policy — import ordering here matches the crate's existing conventions (messages/, etc.); fmt drift is at parity with upstream (156 diffs pre-split → 171 across the same volume).

Verification (matches CI gates)

• cargo check --workspace --locked ✅
• cargo test --workspace --locked ✅ 1,799 tests, 0 failures (tui: 708)
• cargo clippy --workspace --all-targets -D warnings ✅ clean
• 208/208 functions preserved vs. original — 0 missing, 0 extra
• Diff: app.rs deleted, 11 new files, +7,892 / −7,807 (the +85 is module glue)

No behavior change, no test changes, no dependency changes.

[nhogenson]

Closing per request — refile as SlugThugLabs when the org token is available.