chore(test): guard a plan against a single-exercise monoculture, and close #89

[kilianmc]

Closes #89 as a non-defect, and keeps the part that is genuinely owed.

Why #89 closed

It tracked "~36 exercises carry ~80% of every plan" as a library concentration defect. Measured over 144 plans (both disciplines × three grade bands × 1–7 days/week × three weakness settings, crossed with a full-vocabulary and a bouldering-wall-only equipment column):

	full vocabulary	bouldering wall only
distinct rows a plan uses (median)	68 of 94 eligible	34 of 38 eligible
smallest head carrying 80% of blocks	36	18
head ÷ rows actually used	0.55	0.55
max single row, blocks / minutes	11.3% / 13.6%	14.8% / 10.9%

There is no concentrated head: an even spread over the rows a plan uses reads 0.80 and it reads 0.55. The head count is mostly set by the eligible pool, which the climber's equipment fixes, so it moves on a purchase rather than on a defect. Ruling 47 had already measured the one route to moving it — re-weighting the rotation — as moving presence, not proportion, because a 2-session week reads only 2 of 16 ring positions.

A hypothesis worth recording as dead: the head is not floored by progression arithmetic. progressed() takes no history and no plan argument and the dose is a pure function of the week's ordinal, so 68.3% of (block, exercise) pairs appear in exactly one week of their block. Nothing requires a row to repeat.

What ruling 55 keeps

PR #120 fixed a real monoculture at its cause, and nothing pinned the result — no test in the repo asserted a per-plan single-exercise share, so a library or ranking edit could have walked it back silently.

test_no_single_exercise_DOMINATES_a_plan asserts a per-plan ceiling on the single largest exercise over the existing twelve-row sweep × six session counts × three weaknesses = 216 plans: 17% of blocks, 23% of seconds. Both halves, because #120 stated its result in minutes while #89's metric was blocks. OPEN_CLIMBING_KEYS is out of numerator and denominator — those blocks arrive as ruling 27's length fill by ruling 29's decision.

Shown to fail

Narrowing prescribable() to one row per cell — the generalised pre-#120 condition — puts limit_boulders at 17.50% of blocks and 30.78% of minutes, over both ceilings, on all twelve rows. Freezing the rotation instead stays green, which is the useful negative: pool narrowness makes a monoculture, rotation order does not. The worst case in both eras is the narrow-equipment column, so a full-vocabulary-only sweep would have been green for the wrong reason.

Known limit, accepted deliberately: on the blocks half the window is narrow (14.85% today, 17.50% under total collapse, ceiling 17%), so that half fires only on near-total collapse and the detection headroom lives in the minutes half. Tightening it would turn red on a benign re-dose, which is why the measured maxima are narrated in the docstring rather than asserted.

Scope

Tests only. No product code, and no new sweep, plan builder or minutes helper — both arms read generate() output through the file's own _input and _block_seconds. npm run check:server green, 1451 passed.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
climb-trainer	Ready	Preview	Sep 10, 2026 8:00pm UTC