✨ Slice C of #854: ask bounded Elicit forms in the REPL

[taras]

Part of #854. Slice C of five.

Stacked on three merged foundations, all already on main:

• ✨ Slice A of #854: retain a safe Agent permission audit and project Agent conversations #857 — Slice A, retained Agent truth and the projection route.
• ✨ Run spawned document work concurrently with <All> (#855) #856 — <All> / <Spawn> concurrent document work.
• ✨ Slice B of #854: run live Agent turns and permission waits in the REPL #859 — Slice B, live Agent turns and permission waits.

Why

The REPL could run a document and follow its Agent turns, but an <Elicit>
reaching it had nowhere to be asked. A question needs a form: fields, options,
validation, and somewhere bounded to draw a long message.

What changes

Before:

• An <Elicit> in a REPL entry had no way to be presented or answered.

After:

• The REPL asks bounded Elicit forms in its drawer, validates submissions
  through Core's own compiled schema, and restores focus to the answer the
  person actually gave.

How it works

The supported schema language is deliberately closed. readQuestionForm()
parses an accepted subset into a deeply frozen ReplQuestionForm; anything
outside it throws ElicitationProviderError naming the offending keyword and
its JSON path. This is not a general JSON Schema renderer and is not meant to
become one — a form the REPL cannot draw faithfully is refused rather than
approximated.

Refusal happens before anything is published: the whole schema is read
before prepareElicitation() and before the asked counter moves, so a refused
schema asks nobody and leaves pending undefined.

Two conditional shapes are refused because Core would read them differently
from anything this form can draw:

• a tested type other than string, at $.if.properties.<field>.type — Core
  evaluates that type, while the form reduces the condition to a string
  equality;
• an if with no required, at $.if.required — properties does not
  require a property to exist, so Core applies then where the field is
  absent. It is refused rather than defaulted to the tested field.

Invalid submission keeps the same question open, reporting normalized
messages keyed to the field that failed; valid submission resolves once with
exactly the validated object. What judges a submission is Core's compiled
schema through prepareElicitation() + validateParsed() — no new Core API
was needed.

The whole question scrolls, through one bounded viewport. The complete
ordered content — the whole message, the form's description, every field with
its annotation, offered values and editable line, every validation message, and
[submit] — is built in order and then windowed. The two scroll controls,
[history] and [close] stay outside that viewport, because they are how a
person moves it and leaves. Rows outside the window are not described at all,
so they are neither drawn nor pointable, and focus is claimed only by a row the
viewport actually shows.

This is what the review corrected. Capacity used to be the drawer's height less
every form row, which at the minimum accepted 72×20 terminal left
drawer:value:feedback, [submit] and [close] described but never placed —
and a row layout cannot place is not something a person can see or point at.

Editing is scalar-safe: one erase removes one Unicode scalar, not one UTF-16
code unit. The single [history] node sits inside the drawer's modal focus
root while the drawer is open, and is not also drawn beside it.

Focus restoration is causal, not positional. ReplState.restore for an
answer carries { kind: "answered", known, answer } — every answer the history
already held, plus the exact object sent — so the claim resolves to the record
that is new and holds that answer, not the newest one and not one this
answer did not cause. The claim is silent while its row is unprojected and is
spent at the commit where the named control actually takes focus, so later Tab
navigation is not pulled back.

What must stay true

• Nothing about a form is durable. Form values, validation messages and
  scroll position are process-local: they never enter the route and never
  enter the Journal.
• A caller's schema is not mutated. A retained schema stays unfrozen and
  byte-identical after parsing.
• A refused schema is inert. It publishes nothing and asks nobody.

How to verify it

deno task test \
  packages/cli/tests/repl-forms.test.ts \
  packages/cli/tests/repl-journey.test.ts \
  packages/cli/tests/repl-composition.test.ts \
  packages/cli/tests/repl-boundaries.test.ts \
  packages/cli/tests/repl-execution.test.ts
# ok | 29 passed (157 steps) | 0 failed   — exit 0

deno task check                      # exit 0
deno task lint                       # exit 0
git diff --check origin/main...HEAD  # exit 0

The count rose from the reviewed head's 28 / 154 with the correction's new
evidence: the placed-frame walk, the viewport's message walk, two exact-path
refusals and two comparisons against Core's compiled validator.

Three named controls, each applied alone to the final sources and reverted:
overfill-the-drawer reddens the placed-frame row because essential
targetable controls never reach the frame; accept-non-string-condition and
invent-if-required each redden their exact-path refusal and their
semantic comparison.

Eighteen refusal rows assert the exact JSON path of the rejected keyword. Nine
deliberate defects each turned named rows red — accepting an unknown keyword
reddens the unknown-keyword refusals and the publishes-nothing row; keeping
only the first property or enum value reddens both parse rows, five submission
rows and both drawing rows; resolving the provider before local validation
reddens the two keeps-the-question-open rows. They were not re-run for this
delivery: git range-diff reports the rebase as patch-identical, so they
still discriminate exactly what they discriminated when accepted.

Measured weights

Pending on the corrected head 2f0ba006efac10cacb38762e3c1f5b8c972717b4.

No measurement is authorized yet. The two runs attempted at the reviewed head
(36708334659 and 36720937070) were both cancelled and produced no artifact, and
neither describes this head anyway. The earlier successful run 36511769162 at
75662a2b describes the old base and is not reusable. One signed weights
commit follows, with its provenance and the measured floors against the
installed shard counts (deno 15 / node 10 / bun 5), before this PR is merged.

Scope

Included

• Bounded Elicit forms in the REPL drawer, their validation, and focus
  restoration.

Intentionally unchanged

• Slice D Sessions and permission presentation.
• Slice E command, packaged Plan and journeys.
• The terminal mockup and any UI redesign.
• A general JSON Schema renderer. The closed language is the scope.

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.