(WIP) Fix silent agent attrition when spawn positions can't fit goals

[eugenevinitsky]

What

Fixes gigaflow spawn silently dropping agents whose sampled position can't fit a route + goals, and adds tests that fail on any agent attrition.

• spawn_agent: the state commit, route computation, and goal generation now live inside the position rejection loop — a position on short/dead-end lanes is rejected and resampled exactly like a colliding or offroad one. This extends the same retry principle the map goal source already uses (generate_new_goals_from_map's fresh-cell retry, covered by test_map_goal_source_no_attrition).
• generate_new_goals_from_route: returns failure without invalidating/removing the agent; each caller now handles failure explicitly (resample at spawn, remove with a warning at replay init/reset, remove at mid-episode regen as before).
• Init no longer re-rolls goals for gigaflow agents (they already got goals in spawn_agent); the re-roll could fail and strand a removed=1 agent inside the active set.
• test_drive_goals: new test_route_goal_source_no_attrition — on the nuPlan map, all 32 requested agents must spawn live with full goal sets, both at env creation and through the episode-boundary respawn path. Also updates the test_route_goals_full_set_or_removed contract comment to the new caller-owns-removal semantics.
• test_drive_env_smoke: asserts gigaflow spawns every requested agent and that no active agent is ever removed (checked after creation and at every step, which covers the mid-run episode reset).

Why

On the nuPlan test map, 8 of 32 requested agents were silently dropped at env creation (25% batch shrinkage), one agent sat in the active set already removed with re-rolled goals, and after the first episode reset 8 of 24 active agents (33%) spent the whole episode as dead, terminal-masked observation slots. The only traces were [GIGAFLOW ERROR] printfs — no test failed and no metric moved. The failures are position-specific (nuPlan bins are cropped scenario-local maps whose lane graph dead-ends at the crop boundary), so resampling fixes them: with this change all 32 agents spawn and survive on the same map/seed.

Notes

• Before the fix, both new tests fail with active_agent_count == 24, expected 32 (verified against the unfixed header); after it, the full C suite and the Python unit tests (81 passed) are green, and the [GIGAFLOW WARNING/ERROR] spam in the smoke test output is gone.
• Replay mode is unaffected by the spawn change (replay agents don't go through spawn_agent); its total-failure paths keep the previous remove-the-agent semantics, now with explicit call-site handling and a single warning.

🤖 Generated with Claude Code