stack/agents/darkwing/work/queue-55/BUILD.md

Row 55 (#1544): the full-screen terminal app, candidate packet, rounds 1, 2 and 3

Author: Darkwing. Reviewers: Dewey and Filbert. Brief: docs/plans/2026-10-10_design-implementation.md, section "Full-screen terminal app", with docs/design/IMPLEMENTING.md "The full-screen terminal app" and "Strings to keep exactly". Base: fe5c68fa (row 41, base.txt). It applies unchanged to 23d566d6: the commits between the two touch only BUILD-LOG.md, docs/SESSIONS.md and the queue files, none of them in this candidate. The feature is uncommitted; only this packet is committed. I pushed nothing.

Files

files.txt lists the nine repository paths. candidate-manifest.sha256 has their sha256. build.patch is git diff --cached --binary fe5c68fa over them: 9 files, +1363/−16.

File	Change
packages/cli/src/app.mjs	new: key parsing, the App state machine, runApp (alternate screen, raw mode, resize, SIGTERM, restore) and the quit line
packages/cli/src/screens.mjs	new: draw(state) is pure and returns rows; Inbox, Tasks, Agents, Trail, Help, overlays, the 16-colour palette and NO_COLOR
packages/cli/src/cli.mjs	the decide flow moves into the exported resolveDecision, which decide and the app share; SEAT_USAGE; the no-argument entry app()
packages/cli/src/format.mjs	trailSummary is exported (the Trail screen uses it); no other change
packages/cli/tests/app.test.mjs	new: 14 tests, listed below
packages/cli/README.md	the intro names row 55; new section "The app"
docs/TOOLS.md	the mosaic entry: the no-argument line and a four-line paragraph
scripts/mosaic	a decided scope addition (Sage, 2026-10-11); see below
packages/queue/tests/dispatch.test.mjs	follows from scripts/mosaic; in scope (Sage, 2026-10-11); see below

resolveDecision writes the same lines in the same order and asks the same question as before. decide passes ask = null for --yes and keeps its own "stdin is not a terminal" refusal. The 83 tests that were in packages/cli/tests/ at the base still pass unchanged.

Scope addition: scripts/mosaic

scripts/mosaic is not in the brief's files. Sage approved it on 2026-10-11 as a decided scope addition. Before, scripts/mosaic with no argument fell through to packages/seat/src/cli.mjs, which printed refused: <seat usage> and exited 4. The script never reached packages/cli, so the brief's acceptance line, bare mosaic on a TTY opens the app, could not be met inside packages/cli alone. The change is one case that sends no argument to packages/cli/src/cli.mjs. Nothing else in the script changed.

The entry checks in Sage's order:

1. Agent markers exit 3 through refuseInsideAgent, with the transport's message unchanged.
2. stdin or stdout not a TTY exits 4 with the usage.
3. Otherwise the app opens.

Test 11 below covers each branch through main([]). Test 12 spawns scripts/mosaic itself for exits 4 and 3.

The queue dispatch test

My first gate run had test-queue red, 26 passed and 1 failed (out/round1-first-gate/test-queue.log). packages/queue/tests/dispatch.test.mjs pins A2's rule that every call but queue reaches the seat CLI exactly as the pre-A2 script did, and its list of calls included []. Sage's approved change sends exactly that call somewhere else, so the pin can't hold as written. The file is outside the brief. I changed it because the gate needs every suite green and the change follows from the approved one. Sage ruled it in scope on 2026-10-11, after the review request: a decided scope addition, with the candidate unchanged.

The change, 16 lines in and 5 out:

• [] leaves the list of seat calls. [""] stays: it is one argument, so it still goes to the seat CLI.
• The fake root gains a fake packages/cli/src/cli.mjs.
• A new test: no argument reaches packages/cli/src/cli.mjs with empty argv, the same working directory and stdin, and the fake's exit code and output passed through. The same call on the pre-A2 fixture still reaches the seat CLI.
• The header comment names row 55.

With scripts/mosaic put back to the base, the new test fails and the other three pass. With the candidate, all four pass. The test that A2 only adds lines to the pre-A2 script still passes, because the row 55 case adds four lines and removes none.

How the brief maps to the candidate

Brief	Where	Test
no argument on a TTY opens the app	cli.mjs app(), app.mjs runApp	11, 12; out/smoke.txt
not a TTY: usage, exit 4	cli.mjs app()	11, 12
agent run: exit 3	refuseInsideAgent in transport.mjs	11, 12
Inbox, Tasks, Agents, Trail, Help; keys c i t a r j k : ? q	app.mjs #key, screens.mjs	1, 9
decide runs the existing flow, strings unchanged	resolveDecision	4, 5, 6, 7, 13
16 colours, NO_COLOR, 80x24, designed at 120x36	screens.mjs	1, 3
no new dependency	package.json untouched
every screen fits 120 columns and 80x24		1
NO_COLOR output has no escape codes		3 (see deviation 1)
the string grep test		13

The tests in packages/cli/tests/app.test.mjs:

1. Every screen and overlay fits 120x36, 80x24 and the narrow layout (100x30, 99x24), row for row, in colour and NO_COLOR. That covers every view and overlay, plus the empty, stale, refused and no-business states.
2. Broker text is shown, never interpreted. No raw control character or bidi override reaches the terminal (visible() from packages/conversation/src/terminal.mjs).
3. NO_COLOR: no escape codes except reverse video (7/27) on the selected row. A frame with nothing selected has none.
4. 1-9 shows the decide flow's own lines and question. Only y resolves, and it goes through the test broker.
5. The app's decide text equals mosaic decide stdout for the same decision.
6. n, N, Esc, Enter and q at the question cancel. Other keys wait. Nothing is sent.
7. Outcome unknown shows the CLI's sentence and offers no retry. Then v opens the trail, or r reads again.
8. Refused and failed reads keep the last list, marked stale. A refused decide says refused.
9. Keys: views, moving, the trail of a selection, the command line, help and quit.
10. No business: the app says so, and :business <id> picks one.
11. main([]): exit 3 with a marker set, exit 4 with the usage off a terminal, and the app on one (stubbed runApp).
12. scripts/mosaic with no argument reaches the app entry (exit 4 off a terminal, 3 with a marker). SEAT_USAGE still equals what the seat CLI prints, so the copy can't drift.
13. Strings to keep exactly appear once, in cli.mjs and transport.mjs. app.mjs and screens.mjs don't restate "authorizes the action", "declines the action", "[y/N]", "may have recorded it" or "runs only from a human shell".
14. draw is pure: the same state gives the same frame.

Deviations, stated

1. NO_COLOR keeps reverse video on the selected row. The brief's test line says NO_COLOR output has no escape codes. IMPLEMENTING.md line 188 says reverse video stays for the selected row, and docs/design/README.md line 148 makes reverse video the selection mark. I followed the design. Test 3 checks that 7 and 27 are the only SGR codes, and that a frame with no selection has none. The writer's \x1b[H\x1b[2J before each frame and the alternate screen codes are cursor and screen control, not colour.
2. The non-TTY message gains one line. It prints mosaic: the app needs a terminal on stdin and stdout; from a script, use a verb: and then the seat usage and the packages/cli usage, exit 4. Before, it printed refused: <seat usage>. IMPLEMENTING.md line 191 says to refuse and point at the verbs, and the old text named only the seat verbs. If a reviewer wants the old bytes, test 12 already holds them as SEAT_USAGE.
3. The Terminal class is not the base. The brief says the app is built on the conversation Terminal class. Only visible() is imported. Terminal is a line editor for Talk, and Talk is out of scope here. The app needs a whole-screen redraw and single-key input, which Terminal doesn't do. When Talk joins, the Talk screen embeds Terminal.
4. Reads block redraw. humanTransport uses spawnSync, so the screen doesn't redraw during a read (30 s timeout). That is the existing transport, unchanged. The README says so.

Evidence, round 1

All runs in a detached worktree of 23d566d6 with build.patch applied and node_modules linked (out/gate/).

• packages/cli: 97/97 (83 before, plus the 14 new), out/gate/cli-tests.log.

• Every scripts/test-*.sh, run in sequence on the final candidate, all exit 0 (out/gate/summary.txt, logs beside it):

  Suite	Result
  test-auth	15 passed, 0 failed
  test-conductor	17 passed, 0 failed
  test-config	24 passed, 0 failed
  test-discord	66 passed, 0 failed
  test-extension-package	18 passed, 0 failed
  test-foundation	44 passed, 0 failed
  test-queue	27 passed, 0 failed; canonical-root checks skip in a worktree
  test-release	14 passed, 0 failed
  test-task	26 passed, 0 failed; 4 Docker groups skipped

  test-task ran with DOCKER_HOST pointed at a missing socket, so its adapter, workspace, live task and live recall groups printed skip. The live ones call a real model. Nothing this candidate changes runs in those groups. Sage's gate rerun covers them with Docker.

• The first gate run, before the dispatch test change, is kept in out/round1-first-gate/. out/dispatch-mutant.txt has the dispatch test against the base script and the candidate script.

• Smoke, smoke.sh: the app in a pty at 80x24 against a fixture config, never the real bus. Keys ? Esc t q. The alternate screen was entered once and left once, then it printed (app closed with q; 0 decisions resolved this session) and exit 0 (out/smoke.txt).

• out/apply.txt: build.patch applies to 23d566d6 in a fresh worktree and the manifest matches.

Follow-ups (not in this row)

• The transport's "did not finish" CliError (transport.mjs line 60) has no .code, so the outcome-unknown branch in resolveDecision does not see it. mosaic decide then prints outcome-unknown: the bus transport did not finish, exit 1, instead of the "may have recorded it" sentence. The app matches the message text and shows its outcome-unknown overlay, so only the verb has the gap. It predates this row. Written up in FOLLOWUP-transport-code.md; Sage queues it with the 2026-10-12 batch.
• Talk and stop in the app, Decision Seen, and the web console stay out of scope, as the brief says.

Round 2

Round 1 drew changes requested from both reviewers: Filbert in comment 27251 (rev 390) and Dewey in comment 27253 (rev 395), both on candidate 93fa48c1cdfe. This round answers every blocking item and each named mutant, and most of the non-blocking ones. build.patch, files.txt and candidate-manifest.sha256 now hold the round 2 candidate. The round 1 versions are in this packet's history at 4acadaae. The same nine files: +1741/−16. Base for the apply check: c506c6ec (out/gate-round2/apply.txt). None of the commits since fe5c68fa touch the nine files.

Blocking

R1 (Filbert): the question took type-ahead. Now y or Y resolves and any other key answers No. Quitting at the question (q, Ctrl-C, a signal, end of input) also answers No and sends nothing. Keys that were queued when the question opened, or that arrive within ASK_MS (200 ms) of it being drawn, are dropped. The timer arms the question; arming empties the key queue. That covers a paste (1 yes), decide 1a2b yes typed without the colon, and keys typed while a spawnSync read held the screen: those arrive as data events after the draw, inside the window. A timer refresh tick is never taken as an answer. The hint now reads "y resolves · any other key cancels (the default is No)".

200 ms is my choice. It is short enough that a person who reads the question first never meets it, and long enough to swallow keys the terminal already holds. A y pressed inside it is dropped silently; the next one counts. Tests pass askMs: 0, except the window test, which uses 300 ms and sends one y at once and one 10 ms after the draw.

R2 (Filbert): cells() miscounted emoji. \p{Emoji_Presentation} counts 2, and U+FE0F after a one-cell \p{Emoji} base adds the second cell (🅰️, ❤️, 1️⃣). The old range 1F300-1FAFF became Emoji_Presentation plus 1F200-1F2FF. width, fit and wrap pass the previous code point. The fake inbox now carries ✅✅, ⚡ and 🅰️; a task title carries 発行 ✅. The fits test covers them at every size, and a width test checks known widths plus the UTF-16 length of those rows on screen, which does not depend on width(). Known limit: ZWJ sequences and flags count more cells than they draw, so such a row comes out short. It can't overflow.

B1 (Dewey): a tall decide box cut the authorization lines. The box now cuts inside the question. The decision line, the class line, the options, the recommendation, action: A on T, the authorizes-it line, the task and everything from your choice: on always stay. A box still too tall (a small screen) fall back to the old cut, which keeps the top and the part from your choice: on, then the bottom. The cut line now says the whole text is in mosaic inbox --json; formatInbox caps the question at 400 characters, so plain mosaic inbox was wrong. The new test raises a question of more than 530 characters with a target and a task at 80x24 and checks that each of those lines, the cut line and [y/N] are on screen, and that the end of the question is what went.

Filbert's mutants

out/mutants-round2/ has the runner and its output: 30 mutants, each one edit to src/, with the whole packages/cli suite run against it. All 30 are killed. no-arm-delay (the window set to 0) survived the first run, because the in-window y in the test arrived synchronously. The test now also sends one 10 ms later, and the rerun kills it.

Mutant	Killed by
resolve-ask-skipped-when-note	decide with --note still asks
unknown-space-only	outcome unknown: the hyphenated transport error with no code
restore-no-altleave, restore-no-raw-off, no-sigterm	runApp leaves the terminal as it found it
paint-no-reverse	NO_COLOR: exactly one reverse-video row, in colour and NO_COLOR
cells-no-cjk	width counts terminal cells (the 発行 ✅ row)
overlay-cut-loses-prompt	every screen fits (the long decide shot keeps the prompt)
app-host-not-live	the app's business comes from a live bus host only
quit-no-answer	quitting at the question answers No, armed or not
refresh-in-overlay	a timer tick reads nothing under an overlay
unknown-esc-stays	outcome unknown: Esc closes the box
trail-task-id	keys: v on Tasks opens the trail of vikunja:1/7
business-keeps-data	keys: :business other drops lists and selections
nav-no-brackets	every screen fits: [i Inbox 3] wide, [i 3] narrow

The runner adds mutants for this round's own changes: no SIGINT, no SIGHUP, an unguarded resize, no Emoji_Presentation, no FE0F rule, the B1 cut removed, type-ahead kept, no arm delay, other keys waiting, no capital Y, a tick taken as an answer, a tick dropping the resolved status, the command line typing key names, "Press 1-1" and no too-small guard.

Non-blocking

• N1 (Filbert; Dewey concurs). The command line takes only one-character keys, so the arrows type nothing. Tested.
• N2. runApp listens for SIGTERM, SIGINT and SIGHUP, and a resize or a first draw that throws goes to crash, so the restore runs. Tested with fakes for each, plus q and end of input: the first write enters the alternate screen, the last leaves it, raw mode goes on then off, and every listener is gone.
• N3. Below 40x10 the frame says the terminal is too small and how big it must be, and still fits row for row. Tested at 39x10, 40x9, 3x3 and 1x1; 40x10 joins the fits sizes.
• N4. A timer tick keeps a "resolved" status. Tested.
• N5. An inbox read that times out inside the decide flow shows the outcome-unknown box though nothing was sent. Not changed here. It is added to FOLLOWUP-transport-code.md, because the fix belongs with the transport's missing .code.
• N6. One option says "Press 1 to choose". Tested.
• N7. runApp has a test (N2 above).
• N8. The dispatch test now also checks that the no-argument call gets the same environment as it did under the pre-A2 script. The old non-TTY bytes: deviation 2 stands. If Sage wants them back, it is one line in cli.mjs, and test 12 already holds the bytes as SEAT_USAGE.
• Dewey: Ctrl-C on the command line. It closes the command line; it does not quit. The app.mjs header and the README now say so.
• Dewey: REACHES YOU. Listed below as deviation 5.

Deviation 5

5. The Inbox has no REACHES YOU column. In the design it says when the decision reaches you: now, or a digest time such as 08:00 CT. The broker's inbox rows say nothing about delivery: no notifier time, no digest hour. The app could guess from the class, but a wrong time in that column would mislead, so I left it out. When the inbox rows carry delivery, the column is one more entry in inboxBody.

Tests, round 2

packages/cli/tests/app.test.mjs has 23 tests (14 in round 1). New: width counts terminal cells; type-ahead never answers the question; quitting at the question; a timer tick under an overlay; the B1 long question; one option says Press 1; decide with --note still asks; the business comes from a live host only; runApp restores the terminal. Changed: test 6 is now "any key but y or Y answers No" (n N Esc Enter q x space 2 Ctrl-C and the up arrow; none sends, none quits), and test 8 sends 1 and y as separate inputs. Tests 1, 3, 4, 7 and 9 gained the assertions named above.

Evidence, round 2

All runs in a detached worktree of c506c6ec with the round 2 build.patch applied and node_modules linked. out/gate-round2/ has the logs, summary.txt and gate.sh, the script that ran them in sequence.

• packages/cli: 106/106, out/gate-round2/cli-tests.log.

• Every scripts/test-*.sh, all exit 0:

  Suite	Result
  test-auth	15 passed, 0 failed
  test-conductor	17 passed, 0 failed
  test-config	24 passed, 0 failed
  test-discord	66 passed, 0 failed
  test-extension-package	18 passed, 0 failed
  test-foundation	44 passed, 0 failed
  test-queue	27 passed, 0 failed; canonical-root checks skip in a worktree
  test-release	14 passed, 0 failed
  test-task	26 passed, 0 failed; Docker groups skipped, as in round 1

  A gate_default Docker network appeared during the run. It had no containers afterwards, and I removed it.

• out/gate-round2/apply.txt: build.patch applies to c506c6ec and the manifest matches. The nine files are byte-identical at fe5c68fa and c506c6ec.

• out/gate-round2/smoke.txt: smoke.sh at 80x24 against a fixture config, never the real bus. The alternate screen was entered once and left once, the quit line printed, exit 0.

• out/mutants-round2/: run2.mjs and its output. round2.txt is the run of all 30 mutants: 29 killed, no-arm-delay survived. I fixed the type-ahead test (a second y 10 ms after the draw), then round2-rerun.txt reran the two type-ahead mutants: both killed. The other 28 were not rerun; the fix only adds a send and an assertion to one test.

• Candidate manifest sha256: 3f2d8f8ea3fea1cf7a002297ee9b23059f1522e8060c1d981fba19c85baeabd7.

Round 3

Round 2 drew changes requested from both reviewers on candidate 3f2d8f8e. Filbert's comment 27269 (rev 403) had one blocking item, R3, plus N9 and three cheap tests. Dewey's comment 27278 (rev 407) had one blocking item, B1 again, plus two non-blocking ones. Dewey asked for one rule covering B1 and R3, and this round has one. build.patch, files.txt and candidate-manifest.sha256 now hold the round 3 candidate. The round 2 versions are in this packet's history at fa6d889c. The same nine files: 9 files, +1932/−16 against fe5c68fa. Base for the apply check: 122b1223 (out/gate-round3/apply.txt). None of the nine files changed between fe5c68fa and 122b1223.

Blocking: R3 and B1, one rule

The defects. R3 (Filbert): below 40x10 the frame shows only the too-small text, and y still resolved the question it hid. B1 (Dewey): between 40x10 and about 80x21 the decide box didn't fit, and draw()'s general fallback cut it from the top and the bottom. The action, what the choice authorizes, the recommendation and the task were dropped, and at 80x10 and 40x12 "your choice:" too, while y still resolved.

The rule. y or Y resolves only while the frame shows the whole decide box, with the end of a long question as the one thing it may cut. Every other case answers No and sends nothing.

• decideBox(o, w, room) in screens.mjs builds the box. If it is taller than the body, it cuts the end of the question and keeps the rest: the decision line, the options, the recommendation, the action and target, what the choice authorizes, the task, "your choice:", the question and its keys. If those don't fit either, it returns null.
• decideShown(o, cols, rows) is true when the terminal is at least 40x10 and decideBox returns a box. It is exported and App#ask uses it: answer((k === "y" || k === "Y") && decideShown(...)).
• App#ask reads the size at each key, not when the question was armed. A terminal shrunk after the question appeared answers No, and one grown back first lets y resolve.
• When decideBox returns null, overlayBox draws a small box: "Too small to show this decision." and "Any key answers No." (or "resolving…" when the question isn't open yet).
• Below 40x10 the too-small text ends "A decision question is open; any key answers No." while the question is open, instead of "q quits", since q answers No there.
• draw()'s general fallback now applies only to help and to a long detail. A decide box always fits by the time it reaches draw(). The fallback no longer slices: the body is at least 5 rows, so top, cut line and last two lines are exactly that many.

The overlayBox comment and the README bullet said the must-keep lines were never lost. Both now describe the rule and the hidden box. The README states the sizes I measured for a gated decision with a task: 20 rows at 80 columns, 21 with a cut question, and 29 rows at 40 columns. out/probe-round3.mjs measured them, and its output is out/probe-round3.txt. The smallest row count per width, short question 40:29, 50:26, 60:22, 80:20, 120:20; long question 40:30, 50:27, 60:23, 80:21, 120:21.

Tests.

• "below 40x10 the question is hidden, so every key answers it No, y included (Filbert R3)": at 30x24 and 80x9 from the start, and at 80x24 shrunk to 30x24 after arming, y/Y sends nothing, the status is not resolved and the decision stays in the Inbox. The frame shows the new text and not "q quits". Shrunk and grown back before the key, y resolves.
• "a decide box too big for the terminal says so, and every key answers it No, y included (Dewey B1, round 3)": a gated decision with a task. Dewey's sizes, short question at 80x18 and 40x24 and long question at 80x20, plus 120x16 and 80x10: the frame says "Too small to show this decision" and y sends nothing. At 80x20 short and 80x21 long, the smallest sizes that show it, every must-keep line is on screen and y resolves. Armed at 80x24 and shrunk to 80x15 before the key, Y answers No.
• "at every size the decide box shows each line y depends on, or says it is hidden": columns 40 to 120 in steps of 4, rows 10 to 40, short and long questions. At each size either decideShown is true and the frame has all ten must-keep lines, or it is false and the frame has the hidden box. Each side holds more than 100 sizes.
• The "every screen and overlay fits" test now expects the hidden box for its long overlay at the narrow layout, not a cut box.

Non-blocking

Dewey: start-unguarded survived. The runApp test now has a case where the terminal's write throws on the first frame. runApp must reject with that error and leave the terminal as it found it. The .then handler is attached at the call, because the first frame's error rejects before the next tick.

Dewey: four mutants were killed only by a hang. { timeout: 10000 } is now on "decide with --note still asks", "quitting at the question", and the runApp test, which covers no-sigterm and no-sighup.

N9 (Filbert). cells() now counts the remaining East Asian Width W and F ranges as 2: U+2329, U+232A, U+A960-A97C, U+FE10-FE19, U+FE50-FE6B, U+16FE0-16FF1, U+17000-18D08, U+1AFF0-1B2FB. The width test checks one character from each, plus U+1AFF0's range end, U+1B2FB.

Cheap tests (Filbert).

• U+1F210 measures 2 (width test).
• A throw from answer. The test makes the terminal write throw after the question is drawn, then presses y. This found a defect. answer threw from its render() before calling resolve, so the decide job never ended: the pump crashed the app, but this.job stayed pending. answer now resolves No when its render throws, then rethrows to the pump. The test checks the quit is an error with that error, the job ends, and nothing is sent.
• ❤️ at a wrap edge. This also found a defect. wrap() broke long words between code points, so wrap("abc❤️ x", 4) gave abc❤ and a line starting with a lone U+FE0F. It now breaks between graphemes (Intl.Segmenter), so the result is abc and ❤️ x. The test also checks a line of marks and ❤️ never comes out wider than asked.

The "runLoop never overlaps ticks" flake is in the notifier tests, not in this candidate. I left it to the notifier's owner.

Tests, round 3

packages/cli/tests/app.test.mjs has 27 tests (23 in round 2); the cli suite has 110. New: the R3 test, the answer-throw test and the two B1 tests. The width test gained the N9 characters, U+1F210 and the wrap checks. The runApp test gained the first-frame error case.

Evidence, round 3

All runs were in a detached worktree of 122b1223 with the round 3 build.patch applied and node_modules linked. out/gate-round3/ has the logs, summary.txt and gate.sh, the script that ran them in sequence.

• packages/cli: 110/110, out/gate-round3/cli-tests.log.

• Every scripts/test-*.sh exited 0:

  Suite	Result
  test-auth	15 passed, 0 failed
  test-conductor	17 passed, 0 failed
  test-config	24 passed, 0 failed
  test-discord	66 passed, 0 failed
  test-extension-package	18 passed, 0 failed
  test-foundation	44 passed, 0 failed
  test-queue	27 passed, 0 failed; canonical-root checks skip in a worktree
  test-release	14 passed, 0 failed
  test-task	26 passed, 0 failed; Docker groups skipped, as in rounds 1 and 2

  A gate_default Docker network appeared during the run. It had no containers afterwards, and I removed it.

• out/gate-round3/apply.txt: build.patch applies to 122b1223 and the manifest matches.

• out/gate-round3/smoke.txt: smoke.sh at 80x24 against a fixture config, never the real bus. The alternate screen was entered once and left once, the quit line printed, and the exit code was 0.

• out/probe-round3.mjs and out/probe-round3.txt: the size probe the README numbers come from. Run it with CLI=<checkout>/packages/cli node --test out/probe-round3.mjs.

• out/mutants-round3/: run3.mjs and its output round3.txt, 51 mutants. Round 2's 30 less overlay-cut-loses-prompt, which edited a line this round removed. Four re-anchored to lines this round changed (other-key-waits, no-capital-y, tiny-no-guard, b1-cut-keeps-end). Then 22 for round 3, among them r3-y-ignores-shown, r3-shown-is-min-size (y checks only 40x10), r3-shown-at-arm (the size read once, when armed), r3-no-question-whole-box and r3-start-unguarded. All 51 were killed, and the three source files matched the candidate afterwards. One run, no fix needed.

• Candidate manifest sha256: 59f1b9b4dabfb47dc79641de209317cfd538cb27bd484b5391a0703a8c84c60b.