Port the product-code parts of #28 after v0.1.0

[alkor2000]

Tracking the parts of #28 (by @dignfei) that are product code and are scheduled to be ported onto main after v0.1.0 is tagged, so that the v0.1.0 candidate stays the soaked rc8 line. Each will land as a commit authored by the contributor once the port is verified against the current main.

• Locale switch transactionality — e6db42d + a32f348 (locale switch transaction, config-missing switch fix, atomic index marker update). Independently valuable; touches internal/cli, internal/config, internal/index, internal/cognitiontxn, two spec/public/ contracts, and text assets. Port needs a rebase over the rc7–rc8 changes in init.go, config.go, and volumegovernance/facts.go.
• Text-asset manifest caching — c707a1e (parse and validate the embedded manifest once; ReadManifest returns a clone). Performance only; review the "cache the error permanently" choice against the current no-cache comment in textassets/catalog.go.
• Scope Change routing for Volumes — bbf725f. On main the Volumes branch of scopechange.Build never projects Entry edits into the Root pointer, so the "wrong target" write cannot happen today; the remaining value is routing Entry-level postimages to the Code Volume, which is a design change to discuss before porting.
• "Prefer AOCI cognition" prompt text — a6d7af7 (AGENTS template and runtime rules). Model-visible text with golden digests; review wording before porting.

Not scheduled (declined on #28 with reasons): the target-index-first workflow (c70f04b … 0dce021; needs its own design and spec/public/ contract first), single-chunk Overview delivery with the challenge count reduced to one (b6c85c9, e51db03, 6be5c95, 33c7566, f0dd398, 22d2a2a), the SQuotaMidRunes relaxation, and the posture-relaxation auto-authorization (c7786c6, superseded by the blocker partition landed from #27).

[alkor2000]

Correction to the "Scope Change routing for Volumes" bullet above (bbf725f). It says that on main the Volumes branch of scopechange.Build never projects Entry edits into the Root pointer, so the wrong-target write cannot happen. That was wrong. #47 traced it: the planner parses cfg.IndexPath, which under Volumes is the Root manifest with zero Entries, so an Entry candidate with an empty current_entry_sha256 was accepted and projected as an insert into aoci.txt (entry_creates counted it, risk low), and a preview like that would have applied under auto authorization. I reproduced it on a copy of a real Volumes repository.

Resolved in 730007f (v0.1.0-rc12) by refusing Entry content in a Volumes candidate set before projection and by retaining a changed source's old fingerprint instead of blocking the policy-only transaction (details in #47). The routing of Entry-level postimages into the Code Volume that this bullet proposed is therefore not planned: under Volumes the Scope transaction stays policy-only and Entries are written only through Maintain.