Say which rung answered, when the ladder had to relax one

[EnesYilmazcode]

Closes #90.

What changed

The engine has computed relaxed and rung since the ladder shipped and js/app.js read neither. Every rung below asked produces rows that meet the need, and if (meets) strip = '' prints nothing whenever the rows meet the need, so all four relaxations were silent.

• js/engine.js exports RUNGS: the ladder's names in the order it walks them. query() now builds its rung list from RUNGS instead of keeping a second copy inline, so the list the screen is tested against and the list the ladder walks are one list.
• js/state.js gets rungPhrase(rung, { needed, maxWalk }): one plain sentence per relaxed rung, in print and in speech. It lives there rather than in js/app.js because js/app.js touches the DOM at import and cannot be loaded under node, and because js/state.js already holds the row phrasing for exactly that reason.
• js/app.js stores state.rung / state.relaxed off a ladder run in answer(), paintList gives the rung the one strip the list gets, and #say leads with the same sentence so this is not a visual-only disclosure.
• sw.js shell gzip figure re-measured, because the shell grew.

The wording is not the issue's wording, and here is why

The issue proposes "Nothing ordinary is free. These are computer labs and meeting rooms." I measured that sentence and it is false every time it would print.

A rung wins by finding LADDER_QUORUM = three rooms, and the ladder overwrites a held rung whenever a later one reaches quorum. So any-type does not mean no ordinary classroom is free; it means fewer than three were, and the ladder preferred three labs to two classrooms. Same for further: it means fewer than three near you, not none.

Replaying the shipped app's own ranking (rank() then shape()) beside query() over the committed index — 99 origins on a grid around the Oval, Mon–Fri 2026-09-14 to 18, hourly 08:00–20:00, at a 30 and a 60 minute ask, 12,870 answers:

ladder relaxed	9,486 of 12,870
relaxed and the list still had rows	510
further/anywhere/longest over a list with rows: rooms a reader counts as free and long enough inside the walk bound	0, 1 or 2 — never three (85, 7, 16 of 108)
any-type: ordinary classrooms on screen that were free and long enough	1 or 2 — never zero and never three (15 and 151 of 166)

So "Nothing ordinary is free" would have been contradicted by the rows underneath it in all 166 cases. Every sentence therefore counts to the quorum: "Fewer than three ordinary classrooms near you are free for 1h00. Vacant reached into computer labs and departmental rooms." The room words are measured too — over those 166 answers every free, long-enough, non-classroom row on screen carried a secondary type (2P computer teaching lab 130, 5K departmental seminar room 120, of 272).

longest never fired once in the 12,870, so its sentence is the only one with no measurement behind it, and it says the least for that reason.

Strip precedence

Rung first, and only one strip. Both are true at once on 131 of the 510 lists that had rows (25.7%), so printing both would stack a second paragraph over the list a quarter of the time. The row counts describe the rows; the rung describes which question the whole list is answering. The unknown-hours caveat is deliberately untouched — it is not a strip, and a list of rooms nobody publishes hours for has to keep saying so whichever rung found them.

The #say sentence is announced only when the list has rows, because that is the only screen the strip is on: the empty screen already names the same fact in its own words, and a live region reading out a sentence that is nowhere on the page is worse than one that says nothing.

Tests

Eight new tests in scripts/test/engine.test.mjs. The one that matters:

every rung the ladder can return has a sentence, and a new rung fails until it does

Proved in both directions by temporarily adding a rooftop rung: added to RUNGS + the runner, the coverage test fails; added to the runner only, the "walks the names in RUNGS" test fails. Also included: a probe over the committed index asserting that no rung sentence is ever contradicted by the rows printed under it (420 relaxed answers, 19 over a list with rows), plus tests that the walk bound in the sentence is the bound the rows were cut to, that the spoken form says "1 hour" where the strip says "1h00", and a source scan of js/app.js for the wiring and the precedence (the issue's own grep -n "relaxed\|rung" js/app.js returning nothing is what this replaces).

npm test: 674 tests, 670 pass, 0 fail, 4 skipped (baseline on main: 666 / 662 / 0 / 4).

What I did not do

• scripts/shoot.mjs gets a frame of a relaxed answer — not done. shoot.mjs drives real Chrome and there is no Chrome or Chromium in this environment, so I could not capture a frame, and I was not willing to add a frame definition I could not run or verify. Its pinned scenario (Thompson Library steps, 2026-09-16 10:20) also never relaxes: from the Oval the ladder came back on asked in all 250 probes there, so the frame needs a second pinned origin. Concrete cases that do produce one, if you want to shoot them: from 39.9915, -83.0230 — Mon 2026-09-14 08:00 at a 2h ask gives any-type, 10:50 at 30 min gives opens-at, 17:50 at 30 min gives further, each over four or five rows.
• The rung describes the ladder's verdict, not literally the rows. js/app.js builds its list from rank() + shape(), which have no ladder in them, so the strip and the rows are two answers to one question. The quorum-honest wording plus the probe test is what keeps them from contradicting each other, but the clean fix is for the list itself to come from query(). That overlaps The ga flag ships in every room and the ranking never reads it, so 98 departmentally controlled rooms rank as if they were general classrooms #89 and would change which rooms are ranked, so it is not in this PR.
• Cost: the disclosure runs the ladder as a second sweep. Measured under node on the committed index (425 rooms, median of 200 warm runs): rank() 1.02 ms, query() 1.01 ms, so an answer goes from about one millisecond to two.
• sw.js's SHELL_CACHE stamp is left at its current SHA; the script's own docs say the committing job runs scripts/stamp-sw.mjs. Only the gzip figure in the header was corrected.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_017xiX95r6nnUUmhD3KwMnWM

---

Generated by Claude Code