IR-86: size the golden Linux `test` job's start floor from the measured S4 run. Golden 34946493637 consumed ~81.4 decimal GB before cleanup; plus the 32 GiB reserve that is ~115.8 GB, so the Linux test job now admits at 130 GB (~14 GB margin). Docs/end floors stay at 32 GiB and keep their pre-cleanup order. Windows start-floor calibration remains separate IR-86 debt.

Gate (doyle): workflow-comment + one constant change in golden.yml, register + toml only otherwise; parent 84e20ddd (IR-136, now main); `traceable-reqs check` 0.4.1 exit 0 in a throwaway worktree. hertz: shell probes — old 110.7 GB accepts, new refuses; admission/reserve boundaries, df-error and real-disk arms all as expected (.spt/preserved/hertz-ir86-linux-floor/handoff.json on kitsubito).

Co-authored by: hertz

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01SgutL1VJbbSwoa7NSsKuHp

---

Folded (doyle): the Windows half rides the same PR. 0cda6235 raises the Windows golden `test` start floor from 32 GB to 128 GiB, derived from run 34262154550 / job 102182665033 on hfenduleam (primary log hashed at `.spt/preserved/hertz-ir86-windows-floor/`): START 97.6 GB → END 25.3 GB, delta 72.4 GB, plus the 32 GiB reserve = 99.4 GiB, rounded to the next binary boundary with the ~28.6 GiB margin stated as margin. The register records that this floor would have refused hfenduleam before today's retired-target reap (~92 GB free), so owned reclamation is now a prerequisite for that job. `RESOURCE=disk` refusal semantics unchanged: capacity refused, workload untested. Gate: one constant change, `traceable-reqs check` exit 0, hertz's PowerShell guard probes 10/10.