[harshabose]

Implements docs/B1_fix_spec.md: B1 now starts at the AIRBRAKE exit like D and C, and answers one question: the least bus energy as a function of distance, under a stated wing limit.

Stacked PR: the base is feat/c-ardupilot-z-loop (PR #7), so the diff shows only this change. Stack: #2 A, #3 D, #4 Q2, #5 E, #6 D addenda, #7 C, this. Retarget to master once the stack merges. No other branch or PR was touched.

Commits

1. Docs: B1_fix_spec.md.
2. B1 (spec 3-10): collocation.py: make_problem_exit, CollocationOptions.alpha_max, with_options keeps the entry state (it used to rebuild a cruise problem), schedule_summary, bound_check_c, compare through restricted.evaluate_c; bound_check stays as legacy. task_b1 rewritten (per variant: the front over eps_x_m at the 1 m band, the band sweep, the gap to C, the sensitivities). Tables B1_front, B1_gap, B1_sens, B1_schedule. B1Cfg is list-based; an old scalar eps_x_m is rejected with a message. B2 keeps its level-cruise start through _b1_problem_cruise. Configs and READMEs.

Checked against the spec (S-01, real configs)

• 1: cruise entry 10 m band: 6.083 kJ, 83.4 m, 6.09 s, 10.00 m. B2's cruise-start solve: identical arrays and energy (6083.479 J) on the C tip and on this branch.
• 2: exit entry, 10 m band, attached: 6.265 kJ, 95.9 m, 7.41 s, 9.60 m.
• 3: attached, 1 m band, eps_x 300: 8.922 kJ / 72.7 m / 6.69 s (end band 0.5, carry); 8.918 / 72.8 / 6.69 (1.76, carry); 8.921 / 72.6 / 6.68 (0.5, idle).
• 4: attached at eps_x 40 and 57: not converged from three starts; rows kept with energy blank.
• 5: c_envelope at eps_x 40: 8.226 kJ, 38.7 m, 3.84 s (end band 0.5: 8.220); with the 34.4 deg pitch limit 7.748 kJ / 39.6 m / 3.74 s at eps_x 40 and 7.605 / 49.5 / 4.51 at eps_x 300.
• 6: C at the reduced-search optimum: 12.68 kJ, 0.52 m, 40.1 m (as in the spec). B1 c_envelope at that excursion and distance: 8.34 kJ, 34% less.
• 7: replayed peak wing angle: attached 13.4 deg (limit 13.6); c_envelope 30.47 deg (limit 30.5).
• 8: rotors come in at about 2.3 s on S-01 attached, at idle until then.
• 9-11: S-02 has below_vs true; bound_check and compare still run; the old scalar eps_x_m is rejected.
• 175 existing tests pass; ruff, mypy and lint-imports clean. No tolerance exceeded.
• A reduced end-to-end run.py --only B1 (S-01 and S-02, eps_x 40 and 300) writes the four tables and the report section.

Not run

The full 6-scenario, about 100-solve block (2 to 3 hours on one core) was not run; only the reduced study above. No new tests (the spec leaves them to the owner).

Notes for review

• B1.gap_c is an addition. The B1 tasks are independent processes and cannot read C's optimum, so the gap evaluates C at B1.gap_c (paste the four values from C_params), or at the register values with q = 2 if it is null. The table says which.
• The attached gap rows do not converge at C's 40 m (attached cannot go much below 73 m at a 1 m band). That is a solver result, not proven infeasibility.
• schedule_summary takes (prob, check, gate), not just the replay, because it needs the dead zone and the stall speed. t_rotor_on is the first time after the throttle has rested at its minimum that it exceeds idle + 0.05 (the spec does not define idle for a carry start). An extra column alpha_w_peak_above_vs_deg carries check 7's gate.
• The rotor-start flip and the 0.5 m/s end band run on the first variant (attached); the 34.4 deg pitch limit on the first absolute-angle variant (c_envelope).
• Eleven exact-Hessian solves at once ran out of memory (std::bad_alloc) on a 24-core workstation; limit parallel.max_workers if memory is short.
• The old Pareto figure no longer fits a fixed 1 m band and is not produced; B1_schedule holds the replay series for a plot.
• B1 is clairvoyant, open loop and has no rotor lag (C has rotor_lag_s); the end band bounds climb as well as sink (C bounds sink only). VRS is only reported (B-3). B2 is not on this entry yet, so B-9 stays open.
• The frozen tests/data/placeholder_configs/study.yaml B1 block follows the new format.

🤖 Generated with Claude Code