#1194 exact native V1 shared Server feasibility, 2026-10-14

Source PR2721 head: 7770ebb9cb04106c6f0205c12082f748e35d28d1
Normal main merge: c0e9a387d89f4ec2e47bf21326a6f5ab5f0ad38b, 03:35:01Z
Fresh main ancestry verified. No runtime/default/dependency/release changes.
Dedicated Actions run 38506584150: SUCCESS, both native Docker jobs SUCCESS.
Fixture SOURCE_COMMIT (PR test merge): 1e86643ecbd767417cf61d08a1e9ba42bf8c7dd8
1.18.1 job115665384158 / artifact11868630502
1.18.34 job115665384478 / artifact11867354420
10 required guards SUCCESS before ordinary exact-head merge.
Broader QA jobs were still pending; this is not a full-QA PASS declaration.
Subsequent broad run38506584050: server-unit job115666011746 failed in three
unchanged files (legacy sessions.network_id NULL insert, observer task dispatch
403 versus 200/202, due-date filtering extra row). Desktop-push job115666013312
also failed on that observer assertion. Server/runtime source is unchanged from
c103, but no separate exact-base run was performed, so these are pending triage,
not proven stale tests or a proven product regression. Tracked separately in Hub.

Per version, 11 checkpoints passed (22 across the two exact packages):
1 exact binary/version and loopback model environment
2 unauthenticated 401 and authenticated native health/version
3 --pure Server directory header, separate flat MCP/config and vetted tool deny
4 concurrent sessions: actual bound MCP identity A/B, alpha/beta model routing,
  vetted disabled tool names absent from actual model requests
5 missing session 404 and foreign-session read characterized
6 two native full attach TUIs rendered against one Server
7 toast measured from both directory event streams
8 closing exact attach A leaves attach B and Server working
9 persistent host restart retains both exact sessions and verified history
10 offline scratch native import preserves both identities/parents/parts
11 restored scratch session accepts a new synthetic turn

Both versions: B directory GET of A session returns 200, same ID, directory A.
Directory routing is not session authorization. The future bridge must validate
network/node/Server identity/canonical directory/exact session before readiness
and writes; no claim that upstream rejects foreign-session access.
Both versions: the measured directory-scoped toast reached A's event stream,
not B's. This is one synthetic two-directory observation, not universal native
UI/platform notification acceptance. Existing native FIFO/owned-turn rules stay.

TEST-ONLY: Docker non-root, --network none at execution, local synthetic model
and MCP identities, private tmux socket and exact pane/session cleanup. Wrapper
and Linux-x64 package versions/integrities are pinned in the fixture input.
No Hub account, vendor model, real credential or original node/history accessed.
Raw native feasibility does not implement an ANet shared-node bridge, persistent
host launcher, original-node login, production migration or disaster recovery.
The two projects share a synthetic provider credential and equivalent policies;
different provider credentials/policies and same-directory multiple identities
are not admitted by this first implementation contract.

Retained failures: initial 934792f84 run38506238370 stopped at fixture model
argument index; its QA trigger guard failed separately. a416d2a61 run38506378312
passed first five native checkpoints, then the fixture used a session name as
a pane target. 7770 uses exact returned pane IDs, retains dead-pane diagnostics,
and both complete gates passed. No product failure inferred from fixture bugs.
Three obsolete in-flight own-branch CI runs were cancelled, not counted PASS.
No local Docker/native tests were run; static source/shell/Python/JSON and diff
review only. Long native checks ran in Actions.

Decision: proceed to #1195 only within the same trust cohort/unique-directory
contract, with mandatory identity guard and independent host ownership.
#1196 migration/recovery and actual Owner canary remain separate pending work.
