Prove live workspace coverage recovery

This commit is contained in:
Meta Vens 2026-07-16 03:24:54 +01:00
commit 348024071e
12 changed files with 711 additions and 36 deletions

View file

@ -890,8 +890,8 @@ the indexed content. Receipt/head constructors and writers remain projector-
private. Live query eligibility re-observes Qdrant; a stored receipt proves only
that one exact observation completed.
The crate-private execution path now exists but remains deliberately
production-unreachable. It validates plan-sealed text hashes and vector
The execution path is reachable only through the reserved packaged workspace-
coverage projector managed by Idunn. It validates plan-sealed text hashes and vector
dimensions, creates or authenticates one exact managed claim/epoch Qdrant
collection, writes no empty upsert, whole-scrolls typed payloads, rejects cyclic
pagination plus duplicate/extra/missing/substituted points, and terminal-CASes
@ -899,17 +899,15 @@ the receipt/head against the exact current Body authority, immutable plan,
running claim/attempt, and prior coverage head acquired at start. A CAS loser
cannot mint a receipt; failure can terminalize after Body advance.
The service fork is resolved in favor of a dedicated workspace-coverage
projector. Before it becomes reachable, claims must bind an executor
incarnation and startup lifecycle receipt authenticated by the reserved managed
service. The pulse needs an exact `Current` result to avoid endless reprojection
and a sealed text-rematerialization path rather than caller-supplied workspace
bytes. Abandoned-claim recovery may use only a newer latest managed launch and
its correlated ready heartbeat to fence the old incarnation in one Body-store
CAS. Time, Qdrant state, generic policy, process guesses, and caller strings are
not abandonment authority. If Body advanced, recovery terminalizes the stale
claim and lets normal acquisition derive a new plan; it never resurrects stale
work.
The dedicated workspace-coverage projector binds claims to an authenticated
executor incarnation and reserved managed-process launch. Its pulse has an exact
`Current` result and rematerializes text through the authenticated historical
Body session rather than caller-supplied bytes. Abandoned-claim recovery admits
only immutable host-signed termination, one causally linked replacement launch,
and that replacement's latest signed ready heartbeat before one Body-store CAS.
Time, Qdrant state, generic policy, process guesses, and caller strings are not
abandonment authority. If Body advanced, recovery terminalizes the stale claim
and lets normal acquisition derive a new plan; it never resurrects stale work.
### Dedicated service and exact semantic proof
@ -931,13 +929,15 @@ success. Body history derives retirement candidates for terminal non-current
claim collections; current and running collections are preserved, and Qdrant
deletion requires exact managed metadata.
Recovery is deliberately not present. A newer launch and ready heartbeat prove
a successor exists, not that the old writer died. Honest recovery first needs a
supervisor-owned termination observation bound to host+boot incarnation,
process-instance identity, exact old launch/PID/executable/policy, and the old
launch-correlated heartbeat/provider incarnation. The existing process observer
collapses access failure/PID reuse too aggressively and cannot issue that proof.
Timeouts remain inadmissible.
Recovery is present through Idunn's reserved reconciliation path. Launch,
heartbeat, and termination serialize through one per-launch process-evidence
head. Host-signed termination binds host+boot incarnation, exact native process
identity, executable, policy, and launch; it may seal death before heartbeat one
without turning timeout or PID absence into authority. One termination admits
one causal replacement slot. Recovery then requires the replacement's latest
signed ready heartbeat and commits failed history, epoch+1 successor, and an
immutable evidence-digest receipt in one Body CAS. PID reuse, access failure,
indeterminate observation, stale ready replay, and timeout remain inadmissible.
This map must change when ownership changes. Historical scars belong in git,
evidence, or an explicitly archived note—not in the machine's Modeling state.

View file

@ -1925,3 +1925,42 @@ Next: create/run a GUID-scoped live smoke through real packaged supervisor and
projector binaries, local Qdrant/Ollama, exact child kill, causal replacement,
epoch+1 recovery, and restart-time receipt authentication. Restore the prior
Qdrant container state afterward. Reboot/logon remains permission-bound.
## Live packaged workspace coverage recovery proof (2026-07-16)
Feature-gated binary `epiphany-workspace-coverage-recovery-smoke` constructs
GUID-scoped repo/runtime/Body/Verse state, uses the real packaged supervisor and
projector siblings, warms local Ollama, derives its exact fenced Qdrant collection names, launches
through specialized bootstrap, catches the epoch 1 claim, kills the exact child,
runs Idunn reconciliation, verifies epoch 2 authority and fresh-store-reopen
recovery receipt authentication, kills the replacement, and removes only its
exact owned collections. Process cleanup revalidates the signed launch's native
incarnation before issuing any kill.
Passed proof:
`C:\Users\Meta\AppData\Local\Temp\epiphany-workspace-coverage-a577f820-b0e2-4733-b628-529ee3bdb143\proof.json`
Receipt: `2b44f28f-7c0f-4ad2-aefe-d7c8222e8642`. The proof binds executable
SHA-256 `d262b7f49d45f15c31589bdfb591988cabcfa9ba65564d3e61493dc00b072093`,
the exact embedded smoke-source digest, source head, and dirty tracked-source
diff digest. It verifies both owned collections absent before emitting success.
Operator orchestration restored the
local Qdrant container to its prior stopped state; that is not a harness claim.
The live rite exposed two real defects and did not pass until both were cut.
Windows boot identity no longer opens privileged PID 4; it queries native
kernel boot time and a focused test proves availability/stability under the
current normal user. Recovery no longer deadlocks its successor: the exact v3
replacement incarnation may resume its own running claim, while a wrong
incarnation remains contended. The harness uses file-backed supervisor output
because Windows grandchildren can retain captured pipe handles, and it has
exact-incarnation process cleanup and exact-owned-collection cleanup guards.
Verification: five native process-observation tests, six process/recovery tests,
fourteen supervisor tests, fifteen projector tests plus one ignored live test,
and feature-enabled all-target check pass. The separate live Qdrant/Ollama
projection test passed before the recovery smoke.
Next: audit deployability against the swarm readiness plan and live host state.
Do not conflate code-complete physiology with the permission-bound reboot/logon
proof; no reboot without explicit live operator approval.