⚡️ Reuse the layout work REPL typing and focus did not change (#881)

[taras]

Why

Typing in the REPL does far more layout work than a keystroke needs, and the
work grows with the length of the selected entry rather than with what
changed. Measured on the production commit path, one keystroke beside a
222-line selected entry cost 2,247 engine measurements — for 52 terminal
bytes.

Follow-up to #881's accepted parity implementation. Stacked on #887.

What changes

Before: every dispatched action — including every keystroke — built a view and
ran the whole frame preparation before reducing, and read one field of it.
Source fitting was recomputed from scratch on every frame. A focus move
painted the entire frame twice.

After: an interaction that changes nothing a measurement decides reuses the
geometry it already has.

Selected reading	Before, per keystroke	After
12-line entry	127 measurements	5
222-line entry	2,247 measurements	5

Five is the fixed cost a complete frame always pays. It no longer grows with
the entry.

How it works

keystroke → tree.dispatch → readsGeometry? → reduce → commitReplFrame → prepareFrame → prepareReading
                                 │                                                          │
                            (false: no frame measured)                        (inputs unchanged: reuse)

Three independent changes:

1. readsGeometry, exported beside reduceRepl, says which actions need
   the frame's admission. Only the four scrolls do; the reducer reads
   admission in exactly four branches. act() prepares only for those, and
   skips the build() and the screen.size() read for everything else.
2. prepareReading splits into geometry and decoration. fitGeometry
   performs every engine render and answers where each line is cut and how
   wide each piece is; decorate builds rows from the lines as they stand
   now. A ReplFitStore owned by runReplProgram's scope holds the last
   geometry with the inputs it came from.
3. Focus settles before the paint. A traversal has already moved focus in
   the tree, so the loop takes it before composing instead of drawing the old
   control and correcting it.

Review guide

Start with: packages/cli/src/repl/fitting.ts — the geometry/decoration
split and createFitStore.

Then review:

1. readsGeometry in application.ts, and the gate at its call site in act()
2. The store threaded through prepareFrame / commitReplFrame
3. The focus settle in act() and the first-frame paint
4. The regressions, especially RS2 and RS3

Look carefully at:

• Invalidation: text, depth, room and terminal size each invalidate
  independently, and inputs are compared rather than digested.
• Decoration runs every frame on a hit, so live lifecycle, head state,
  selection and inspected stay fresh without a Journal append.

What must stay true

• Measure → admit → reconcile → draw stays authoritative when geometry
  changes — commitReplFrame is untouched; the store invalidates on size
  and room. Checked by TL3 and the resize journeys.
• Session readings stay coherent — session.reading() is still called
  every frame. This caches presentation work only, never the stale-reading
  shortcut rejected in Navigate recorded History positions with a measured REPL rail (#881, 3/3) #887's review.
• Freedom owns focus — the move is still the tree's; the loop stops
  lagging it. Checked by RS1.

How to verify it

• RS2 proves a complete frame after a keystroke costs the same for a
  12-line and a 222-line reading, and fails if production stops threading the
  store.
• RS3 proves typing, erasing, Tab and Backtab each read the terminal size
  exactly four times through the real loop, and fails if the gate is bypassed
  at its call site.
• S5 proves a 222-line reading is measured once and not again, and that
  depth, room, terminal size and one line's text each invalidate on their own
  — each compared against a freshly measured fit.
• EU2 proves no non-scroll action's answer depends on the admission, over
  all 24 action kinds.

deno task test packages/cli/tests/repl-source-reading.test.ts \
  packages/cli/tests/repl-presentation-style.test.ts \
  packages/cli/tests/repl-layout-admission.test.ts \
  packages/cli/tests/repl-entries.test.ts packages/cli/tests/repl-forms.test.ts
deno task test packages/cli/tests/repl-journey.test.ts \
  packages/cli/tests/repl-agent-journey.test.ts packages/cli/tests/repl-terminal.test.ts

→ 52 passed (218 steps) and 56 passed (146 steps).

Seven mutation controls, each required to break the test it names, are in
.reviewer/issue-881/pr4-evidence/.controls.py.

Scope

Included

• The action gate, the fitting split and store, and the focus settle.
• Regressions for each, and a counting renderer in the test fixtures.

Intentionally unchanged

• The History rail's own measurement — it is one batched render and did not
  appear in the remaining cost.
• The reading tree, which readingOf(view) still rebuilds every frame. That
  is where live phases come from, so retaining it is a different and riskier
  invalidation problem.
• No persistence, Journal, dependency, public API or drawer changes.

New abstractions

• ReplFitStore and ReplRowCut exist because the geometry has to be held
  somewhere with the inputs it was computed from, and because separating the
  measured half from the live half is what makes it reusable at all.
• Each new abstraction has multiple concrete uses or a clear justification.
• No speculative functionality is included.

Risks and limitations

• A keystroke beside a 222-line entry still costs ~1.1 ms of pure JS: the
  reading tree is rebuilt and decorated per frame. No engine calls, but it
  still grows with entry length.
• J2's drag is armed on an absolute count of terminal size reads, and this
  change removes one per keystroke. Its constant moved from 4 to 3, confirmed
  against the instrumented call sequence — too-small-check → drawnAt → measurement → revalidation — rather than against the test going green.
  Every assertion in that test is unchanged.

Scope confirmation

• Every changed file supports the purpose described above.
• Unrelated cleanup and formatting changes are excluded.
• Generated or mechanical changes are clearly identified.
• The description matches the final diff and test results.