EngineDesign overnight sweep: solver window, config-driven stability, flight dedupe, lockout heartbeat, UI cleanup

[Carlsaurus]

Overnight sweep of EngineDesign: Layer 1, the flight/propellant optimizer, the stability model, and the frontend. Stacked on #56 and #57 (their commits are included here; merge those first or merge this one and close them).

Physics that was wrong

• Chamber-pressure solver returned negative thrust on the shipped default config. The supply−demand residual is not monotonic in Pc: negative below ~20 psi (unchoked, efficiency collapses), positive through the operating range, negative again once the injector drop gets small. With Pc_min = 1 bar, Brent converged to the spurious 20 psi root — configs/default.yaml evaluated to F = −333 N at its own configured tank pressures (815/677 psi). The hardcoded 15% "feed loss margin" ceiling also excluded real soft-injector operating points, and the Python solver's endpoint-only sign test reported "no solution" at 550/650/900 psi where the numba kernel found the root. Fix: choked-flow floor (2 atm), tank-pressure ceiling (min tank − 2%), and a highest-Pc bracket scan, mirrored verbatim in both paths. Both now agree at every pressure; canonical results unchanged (0 diffs in a 6-combo snapshot).
• Stability model was methalox with the serial numbers filed off. Fuel latent heat/boiling-point fallbacks were methane's for every fuel; film-gas cp was a fixed 2200 J/kg·K (now CEA γR/(γ−1)); feed-line length for the chug inertance was a hardcoded 0.305 m with no schema field; nozzle-entrance Mach was 0.2 (now solved from the contraction ratio); comprehensive_stability_analysis looked up feed_system["lox"].length — neither the key nor the attribute ever existed — so feed acoustics always used 1.0 m × 10 mm; the legacy and rich paths used different chamber lengths (1L mode 1772 Hz vs 1284 Hz for the same chamber). New StabilityConfig block (n, χ, Mach, damping fractions, regulator, gate allowance) and feed_system.<side>.length, all editable in the Configuration tab. Chug gate centred at gain margin 1.0 (it was 0.80, which called an unstable loop "stable"). Removed: stability_index floors, the water-hammer→margin piecewise map (one branch was literally a constant), and stability/enhanced.py (pintle spatial model fed hardcoded 50/30 m/s velocities; every output was overwritten).
• Ea mixture-ratio step in reaction_chemistry smoothed (20% jump at MR 1.5); the dead reaction-progress computation on every chamber solve removed.
• Layer 1 now refuses a target O/F outside the propellant's CEA table (a propellant switch keeps the previous target; the optimizer used to run against a table that could not evaluate it and returned objective 7317 with nothing to say why), names the hard constraint that held a run infeasible instead of "objective inf / validation failed", and no longer caps chamber length at 1 m. Canonical pintle design_MR 2.55 → 1.4 (ethalox table is [1.0, 2.5]).

Flight

• Optimizer Layer 4 is now the shared FlightSimulation component. The 1100-line copy it replaced hardcoded RP-1 density and an 85% fill factor client-side, so every non-kerolox design capped its fuel load against the wrong tank.
• explicit_capacity_kg in the flight diagnostics carried a bool; rocket/environment fallbacks derived from the request models (the literal copy said propulsion dry mass 24 kg while its own component defaults summed to 16).
• tests/test_flight_propellant_iteration.py read a config out of ~/Downloads; rebuilt on configs/default.yaml, and the tank-cap test now constructs its premise. That test also caught the bool bug.

The lockout ("Take Design 1 before saving — your checkout has lapsed")

Reproduced on an isolated server: hold the checkout, do nothing for five minutes, the next autosave returns 423. The server frees a checkout with no heartbeat inside lock_ttl (300 s); the only heartbeat was an autosave, and autosave writes only when the design changed; the hook also released the token on visibilitychange: hidden, so tabbing away mid-optimization made the run's result write 423. useCheckout now re-takes every 60 s while held and only releases on pagehide.

UI

• Emoji removed everywhere (118 sites).
• Legend and axis-title overlap fixed on every chart with the pattern (Layer 1 convergence, pressure curves, controller, custom plotter, heat flux, flight): legend on top, real bottom margin.
• Design Requirements input caps that assumed one rocket removed (100 kN, O/F 1.5–4.0, 60 s, 3 m…); the propellant's CEA O/F range is shown under the target and flags an out-of-table value; stale stability prose (0.75/0.4 thresholds the code never used) replaced.
• Time-Series simple profile seeds from the design's tank pressures instead of 750/600 psi.
• Config editor: Stability Model section; unreachable Coaxial label dropped.

Schema hygiene

Removed fields nothing read (hot_gas_cp, yield_strength, youngs_modulus, mixing_model) plus the stale hot_gas_cp keys in shipped configs; pydantic ConfigDict; fuel-tank descriptions no longer say RP-1.

Verification

• Python: 433 passed, 81 skipped, 0 failed (base branch: 428 passed, 1 failed). Strict accelerator gate 44 passed; live A/B parity 16/16.
• Frontend: tsc clean, vitest 4/4, production build OK. ESLint is the pre-existing non-blocking backlog; no new findings on touched lines.
• Forward-mode snapshot across all 6 injector × propellant combos before/after every physics change (diffs explained in the session; solver-window and later changes: 0 diffs).
• Short real Layer 1 runs for all 6 combos + a stale-O/F case: no crashes; pintle combos converge in 3 iterations, impinging methalox/ethalox in 30.
• Driven through the UI on an isolated server: forward evaluate at the previously-crashing pressures, time series, flight optimize (clean "infeasible: reaches only 1747 m" rather than a blank page), config tab.

Not verified visually: the chart overlap fix could not be screenshotted while the browser pane was hidden; it is a deterministic recharts layout change (legend verticalAlign="top", bottom margin 20).

🤖 Generated with Claude Code

[Carlsaurus]

Superseded by #57. Three of the five commits here were already on that branch; the two that were not (the overnight sweep and the error boundary) are cherry-picked onto it, so #57 now carries this work as well. Closing so there is one branch to review and merge.