ci/per job cache suffix and pages enablement - #5

…nablement

ci/per job cache suffix and pages enablement

…nablement

ci/per job cache suffix and pages enablement

… (ROADMAP §9 PR R3.2 / ADR-0005)

Move the play-mode shortcut-retargeting key list off the private
`cli/__init__.py:_PLAY_RETARGETED_KEYS` constant onto
`SimulationCfg.play_retargeted_keys()` (static method on the domain
config). The CLI's `env.` → `play_env.` retarget loop in play mode
now calls the method; the set of play-retargetable simulation override
paths lives next to the SimulationCfg fields the --vis / --gpu /
--steps / --dt shortcuts target.

Returns the four paths verbatim (`env.simulation.{vis,gpu,steps,dt}`)
so the CLI use-site is unchanged except for the constant → method swap
— the lowest-risk faithful replacement per ADR-0005's framing. All
four are real SimulationCfg fields (test_configs.py asserts this).

Completes ADR-0005 / R3 (R3.1 EvalCallbackCfg.from_args shipped in
PR #90; this is R3.2).

New `tests/test_configs.py` (3 tests): exact ordered set, every key
names a real SimulationCfg field, and class/instance call equivalence.

Reference audit (per CLAUDE.md §9.1 rule 1): `_PLAY_RETARGETED_KEYS`
had a single use site (the play-mode retarget loop); removed.
`SimulationCfg` added to the existing module-top
`from genelab.configs import …` line in cli/__init__.py (cli → configs
is the allowed layering direction).

Verified: ruff ✓, pyright 0/0/0, full suite 395 passed (was 392; +3
from test_configs.py), `genelab play --help` snapshots byte-identical
(R0.1 gate), configs.py stays torch-free at import (invariant #5),
lint-imports baseline unchanged at 2 kept / 2 broken.

cli/__init__.py LoC: 1051 (refactor start) → 1020 after R3.1 + R3.2
(−31). ADR-0005 §10 estimated ≥35; the actual parse logic was a touch
smaller. Both parsers (eval-callback args, play-retarget keys) are now
domain-owned, which was the substantive goal.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds eight optional SimulationCfg fields mapping to gs.options.RigidOptions:
- contact: enable_self_collision, enable_joint_limit, max_collision_pairs
- solver: solver_iterations (→iterations), ls_iterations, solver_tolerance
  (→tolerance), integrator
- constraint stiffness/damping: constraint_timeconst

All default None ("use Genesis default"). SimulationCfg.rigid_options_kwargs()
maps the set fields to RigidOptions parameter names (genesis-free — keeps
configs.py free of the gs import per invariant #5; integrator stays a string).
InteractiveScene resolves the integrator string → gs.integrator.<name> and passes
rigid_options to gs.Scene ONLY when at least one field is set — so an untouched
config keeps the historic gs.Scene call byte-for-byte (zero regression).

CCD is not exposed (Genesis RigidOptions has no continuous-collision-detection
knob; verified against the installed genesis options surface).

Tests: tests/test_configs.py +3 (empty-by-default, full mapping to Genesis names,
None-skip-but-keep-False). The mapping is unit-tested without a Genesis runtime;
the scene wiring is exercised by the existing genesis_runtime-gated scene tests.

Gates: ruff ✓, ruff format ✓, pyright 0/0/0, lint-imports 6 kept / 0 broken,
optional-dep boundary still green (configs torch/genesis-free), full suite 439 passed (+3).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>