Narrowed residual from GH #177 (whose SpillMode half was resolved by PR #226 — Disk-by-default for host-backed kernels confirmed correct, forcing matrix pinned by test + documented in EMBEDDING.md).
The remaining half, verbatim concern: the host capability feature's /proc, /etc reads bypass the VFS by design. That's fine while 'read-only' means 'kaish doesn't modify the host' — but if a kernel is ever marketed as 'no host observation' (e.g. kaibo's positioning), those reads are a side channel and would need a runtime gate too, not just the compile-time feature axis.
This is a positioning/design decision, not a bug: decide what 'read-only' promises, document it (EMBEDDING.md + README capability-axis blurbs), and add a runtime gate only if the stronger promise is wanted.
Context: split out of #177 during the 2026-07-17 burndown so the closed issue stays accurate.
Narrowed residual from GH #177 (whose SpillMode half was resolved by PR #226 — Disk-by-default for host-backed kernels confirmed correct, forcing matrix pinned by test + documented in EMBEDDING.md).
The remaining half, verbatim concern: the
hostcapability feature's/proc,/etcreads bypass the VFS by design. That's fine while 'read-only' means 'kaish doesn't modify the host' — but if a kernel is ever marketed as 'no host observation' (e.g. kaibo's positioning), those reads are a side channel and would need a runtime gate too, not just the compile-time feature axis.This is a positioning/design decision, not a bug: decide what 'read-only' promises, document it (EMBEDDING.md + README capability-axis blurbs), and add a runtime gate only if the stronger promise is wanted.
Context: split out of #177 during the 2026-07-17 burndown so the closed issue stays accurate.