packaging: pre-carve go/no-go runbook

[mfw78]

What

Adds docs/design/pre-carve-runbook.md: the go/no-go checklist asserted before the physical three-repo carve (nexum-runtime L1, videre-nexum-module L2, shepherd L3). Checks off the runtime-venue-agnostic and cow-on-generic-seam cut gates, the WIT reshape completeness, and additive-extensibility of videre:* against the M4 tip; leaves the classification.toml comment reword and the section 6 carve-time sign-off open.

Why

Milestone M5 requires a single asserted runbook proving both cut gates are green before any git-filter-repo extraction starts, per the 2026-07-15 decision de-gating the second venue from the cut.

Testing

Docs-only change; no code paths affected.

AI Assistance

Drafted with AI assistance from the repo state and the design-notes plan; reviewed by a maintainer.

Closes #406

[lgahdl]

Two checked boxes in section 1 assert gate closure the underlying CI scripts don't actually deliver, and I verified this directly against scripts/check-venue-agnostic.sh at this PR's own commit, not just from memory of the earlier reviews:

• "No charter symbol in crates/nexum-runtime/src; symbol set aligned in PR #449." Checked the script: charter='nexum:intent|value-flow|VenueAdapter|synthesize_venue|nexum:adapter|PoolRouter' still checks only pre-rename dead names (nexum:intent, nexum:adapter, PoolRouter were all renamed away in #428/#443/#447). The current post-rename vocabulary that would actually indicate a leak — videre:types, videre:venue, VenueRegistry, VenueActor, VenueId — still isn't in the pattern (flagged on #449's own review, unaddressed since).

• "Privileged router field deleted... the script refuses any return of pool_router/venue_registry to the runtime." Checked the script: the pattern is literally venue_registry|pool_router, but the actual field in videre-host/src/lib.rs is registry: Arc<VenueRegistry> — snake_case venue_registry never matches that. This scan can't fail on any commit, past or future (flagged on #450's review, unaddressed since).

Both of these are checks that would report a clean pass regardless of whether the regression they're named for actually happened — so checking these two boxes as verified gate-closure overstates what CI actually confirmed. Given this runbook gates an irreversible git-filter-repo extraction, worth either fixing both patterns before carve-time sign-off (section 6), or explicitly re-scoping these two checklist items to "symbol renamed away, script not yet updated to match" rather than "aligned"/"refuses."

Everything else I could cross-check against PRs I've reviewed in this train (the WIT-tree completeness claims in section 3, the additive-extensibility claims in section 4, the cow-orderbook-only gate items in section 2 besides the explicitly-left-open ComposableCoW-comment trip) matches what I found in those reviews.