Fix srv1_floor_memory_budget: use real runner cgroup cap (8 GiB) instead of whole-run-peak × margin

[briansrls]

Summary

• srv1_floor_memory_budget() now returns gunbc_per_runner_memory_cap (the real 8 GiB cgroup cap) directly, replacing within_run_spawn_budget() which computed whole-run-peak × (6/5) ≈ 13.9 GiB — a budget that never matched the actual binding constraint
• Delete within_run_spawn_budget() and the orphaned margin constants (gunbc_ci_floor_runner_margin_num/den, gunbc_ci_floor_whole_run_peak_rss_bytes) which had no remaining consumer after this change
• Rename witness_spawn_width_budget_deconflated → witness_spawn_width_budget_is_per_runner_cap to reflect the corrected invariant

Why the old budget was wrong

within_run_spawn_budget() used gunbc_ci_floor_whole_run_peak_rss_bytes() × (6/5) as the budget passed to memory_aware_spawn_width. That whole-run peak (~11.6 GiB) × 1.2 ≈ 13.9 GiB — larger than the actual 8 GiB cgroup cap. The spawn-width formula therefore operated against a phantom budget, making spawn_width = floor(13.9 × 4/5 / per_shard) instead of the correct floor(8 × 4/5 / per_shard).

Consequence named up front

With the corrected budget and the unmodified per_shard value (2.1 GiB from main), spawn_width = floor(8 × 4/5 / 2.1) = 3. The real cgroup footprint per shard (~3.5 GiB including rustc children from emit_host_gate) means width=3 still exceeds 8 GiB — the floor will remain red on OOM until the footprint reductions land (tidy-wren-707 resolve env de-merge + emit-host batch isolation). That red is accepted. This PR only fixes the budget model; the per_shard reground is a separate PR once real footprint drops.

Not in this PR

• Per_shard reground (2.1 → 3.5 GiB) — operator decision: do not cement the inflated value, recover headroom by reducing real footprint
• graph-release two-phase / malloc_trim — symptom mitigations from Fix CI floor OOM: route srv1_floor_memory_budget to real 8 GiB cgroup cap, raise per_shard to 3.5 GiB #5831, superseded by tidy-wren-707 root fix
• Pre-existing .rs fmt/clippy fixes — separate concern; no .rs files changed here

Test plan

• .dag compile clean: ci_budget_tree.dag and ci_budget_tree_witness_test.dag compile against unchanged ci_floor_measurement.dag (orphaned symbols remain in measurement.dag until a future cleanup)
• Witness witness_spawn_width_budget_is_per_runner_cap passes: srv1_floor_memory_budget() == gunbc_per_runner_memory_cap > 0 < srv1_host_ram()
• within_run_spawn_budget reference removed from all consumers (only consumer was srv1_floor_memory_budget)

🤖 Generated with Claude Code