feat(planner): progress the dose weekly, habituate a beginner's BASE, and put easy aerobic work on the wall

[kilianmc]

What and why

PR 2 of the doctrine-audit tail, at v8.12.0: issues #117 and #118 plus guard 5, ruling 41's authoring, and F11 closed as a declared divergence. Rulings 35–38 and 41 in ct_plan_doctrine.md; rulings 46–48 were taken while building it, and three amendments record what measured differently from the rulings that ordered the work. Server-only — no web/, no migration, no codegen.

piece	measured
#117 week-to-week progression	dose cells identical across a week-pair 78.9% → 60.6%; 0 of 1192 cells carrying a sourced rule remain identical
#118 + guard 5	BASE occurrences 0 → 70, grade gap 1–2 0 → 28, plans drawing it in BASE 0/60 → 40/60, BASE finger blocks that are max hangs/repeaters 115/115 → 0/115
ruling 41 easy_climbing_flush	An Cap weeks with no aerobic work 39.4% → 12.5% (strength 61.1% → 2.8%, power 77.5% → 9.8%)
F11	closed as declared; ruling 38's instruction refused by measurement

Plan digest f95ff03297ba02db → dd29d23ad1962160. Every weekly ceiling holds: hard-energy days >3 outside the taper 0 → 0; An Cap sessions over their phase ceiling 0 → 0. Breadth up, 73.9 → 74.9 distinct exercises per plan, min 52 → 54.

Three findings worth reading before the diff

Ruling 41's row pays into F14 instead of deepening it. general_strength +144 blocks in strength and +120 in power, finger_strength +114 in strength; strength-aspect share of strength minutes 3.87% → 6.14%. A 10-minute wall block fills less of a session than the technique/power blocks it displaces, so the supplementary pass gets more room. This is the opposite of the rejected _try_supplementary override, which cost general_strength 57% — an on-wall row and an off-wall override are not two routes to one outcome.

Ruling 38's instruction was refused by measurement. Seven readings against the whole suite: cap 5 → 14 red, cap 6 → 9 red, cap 8 → co-occurrence red weeks 26 → 128, uncapped → 9 red including ruling 30's out-training invariant, two dense weightings → 8 red each, proportional-to-rank → 5 red. Only the current cap is green. The cause no ruling had: _rotated_pool enters at spread % len(ring), so a 2-session week reads 2 of 16 positions at an offset set by the week number — re-weighting moves presence, not proportion, and presence at 2–3 sessions is what rulings 23/24/30/35 pin. Per ruling 38's own instruction the number was brought back. Closed as a declared divergence (ruling 47); the _spread/_rotated_pool route was deliberately not filed.

PR 2 hollowed one of its own guards, and it stayed green. Ruling 41's row re-based #117's shorter-rest arm from 57 cells to 15, and boulders_on_the_two_minute — 30 of the original 57 — now contributes none, so the arm reads short_rest_boulder_sets alone. Floors re-based 51 → 13 and 448 → 394 with the cause recorded. Recovering the coverage is tied to F26 (ruling 48).

A harness trap, recorded because it nearly cost a wrong conclusion

F11's first measurement round returned all six readings byte-identical — which reads exactly like the five genuinely inert mechanisms already measured in this area. import server.domain.planner.generate as G binds the function re-exported by planner/__init__.py, not the module, so the patch was inert and the harness was measuring nothing. Sabotage the harness and confirm it goes red before trusting a null result here.

CLAUDE.md

• Section covering this change: Findings inbox (empty, and deliberately left so — both new findings have homes in issuesplan.md and ct_plan_doctrine.md, and duplicating them as prose is what that section exists to prevent)
• Confirmed still true, unchanged
• Updated in this PR

Comment budget

• Baseline allowlist entries cleared: 0
• BASELINE_RATCHET in tests/test_comment_budget.py lowered to match

Four new [[exception]] rows and one limit raised, each carrying a measured negative rather than a restatement — including why the cheaper _wall_picks-only aliasing fix must not be re-attempted. Two comments went over their caps during review and were trimmed rather than allowlisted, per the guard's own message that an exception is not a licence to keep going.

Checks

• npm run check green, run once at the end — 1313 passed, ruff and mypy clean
• Any new guard test was shown to fail before it was trusted

Reds captured with production reverted, then restored and re-confirmed green:

• traverse_intervals is dosed (5, 40, None, 100) in week 5 and (5, 40, None, 100) in week 6 of the same strength block; its longer_work rule moved nothing.
• rest_between_sets_seconds fell 100 s -> 90 s. An Cap progresses by harder or longer work and never by less rest.
• (base, sport, endurance) is a pool of 7 and session 0 draws index {0} in all 3 loading weeks: the week term cancelled out.
• a 6a beginner at 5x, gap 1 is prescribed ['hangboard_repeaters', 'max_hangs_20mm', 'no_hang_edge_lifts'] in BASE.
• a 6c 5x plan holds ['hangboard_density_hangs', 'self_resisted_finger_isometrics'] as its whole BASE finger content at gap 1.
• 22 strength week(s) at 5x a week train anaerobic capacity with no aerobic capacity anywhere in the week
• MAX_WALL_TURNS is 10, not 4 · the phases whose wall ring is perfectly flat are (), not ('strength', 'taper')
• easy_climbing_flush is ruling 27's fill at 1200 s and spent 2 sets, not the 1 a 30-minute chunk implies

Prose corrections, none of them catchable by any guard

A comment cited a rest_seconds column CHECK that does not exist (prescription_template checks sets, intensity_pct and target_rpe only — the one-second floor is ours); "these eleven rows" was fourteen; "BASE authors two HOLD rows" — it authors seven, two of them finger; and easy_climbing_flush's user-facing copy said "ten unhurried minutes" while the row is expandable to 1200 s, so it now says "a few" — the same call PR #125 made for lead_endurance_pyramid. Two further false docstrings were corrected by the ranking work, and one test docstring claimed its shape arm "cannot become a hole for a padded authored block" when the rest-free endurance row reached it.

🤖 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 6, 2026 3:59pm UTC