NativeStack2604 verified E2E: sanitized evidence, receipt and decision record (38% readiness before fixes)

docs/decisions/2026-10-04-ns2604-verified-e2e.md

@@ -0,0 +1,121 @@
+# NS2604 verified foundation E2E publication
+
+Date: 2026-10-04. Scope: sanitized historical repository evidence.
+
+## Context and decision
+
+The user asked to publish the verified new-host E2E so future sessions can recover
+its status and fixes, without raw conversations, credentials, local user identities
+or concrete home/profile paths. The machine label NativeStack2604 is retained
+consistently. This bounded job leaves changes in the working tree and a commit
+message for the coordinator; manifests/evidence.json remains coordinator-owned.
+
+Publish the 15-unit, 80-slot status chain as dated evidence. Use the independent
+reviews and adjudications rather than treating executor labels or installation
+logs as final readiness. The north-star action is to make the native foundation's
+remaining work recoverable for complex systems, US-equities research and
+historical simulation.
+
+The aggregation source is the supplied e2e_merge.py and review-prompt.txt, with
+source-file hashes retained in the
+[receipt](../../evidence/receipts/ns2604-e2e-20261004.json).
+Repository format and claim boundaries follow
+[the acceptance-evidence policy](../acceptance-evidence-policy.md) and the recent
+[RTK qualification receipt](../../evidence/receipts/rtk-051-qualification-20261004.json).
+The publication base is native-agent-stack revision
+5358bb564fc42edb54a1edf7b848bf07d7f0997a.
+
+## Method and result
+
+Claude Sonnet executors recorded native host observations. Independent GPT Sol
+reviewers judged one unit each; 22 Claude Opus adjudications resolved all 22
+executor/reviewer disagreements: 14 kept the reviewer's status and 8 took the
+executor's. The merge uses the schema-checked executor
+labels, preserves each independent review, and lets adjudication override the
+final label. Agreement still describes executor versus reviewer.
+
+Recorded executor model metadata names claude-sonnet-5-5. All 15 reviewer result
+records name cx/gpt-6.1-sol-max at requested_effort=max. Workflow runs
+wf_8bf450a9-d1e (17 agents) and wf_cdbf70a0-02c (5 agents) identify the adjudicator
+as claude-opus-5-5, with effort max and agent type evidence-reviewer requested by
+the workflow script. Their starts were 2026-10-04T20:06:45Z and
+2026-10-04T20:08:29Z; completions were 2026-10-04T20:20:30Z and
+2026-10-04T20:19:31Z, respectively. The prompt template as run ships in this PR. The
+[method](../../evidence/artifacts/ns2604-e2e-20261004/method.md) distinguishes these
+provenance levels and dates.
+
+| Final status | Slots |
+| --- | ---: |
+| READY | 18 |
+| PARTIAL | 31 |
+| FAIL | 2 |
+| BY_DESIGN | 12 |
+| INTERIM | 17 |
+| UNJUDGED | 0 |
+| Total | 80 |
+
+**(READY + BY_DESIGN) / 80 = (18 + 12) / 80 = 37.5%, rounded half up to 38%.**
+Every slot remains in the denominator. BY_DESIGN counts an intentional exclusion;
+INTERIM is held by design and contributes zero to readiness. The
+[summary](../../evidence/artifacts/ns2604-e2e-20261004/summary.json) retains the
+per-layer table, and [slots](../../evidence/artifacts/ns2604-e2e-20261004/slots.json)
+retain the individual status chains, bounded reasons, fixes and citations.
+
+This is native executor observation, independent model review and adjudication,
+not an upstream benchmark or a fresh run by the publisher. Host state is as of
+**2026-10-04 19:10–20:10Z**; the final merge is stamped 20:26:16Z. Fixes are pending.
+
+## Fix plan and handoff
+
+The current fix plan is **25 repository-plan and 8 host follow-ups** over the
+33 open PARTIAL/FAIL slots after adjudication. The earlier 30 plan / 11 host
+figure covered 41 reviewer-stage slots before adjudication.
+Keep the 17 INTERIM slots held until their
+existing gates are satisfied, then run their native acceptance and session-use
+checks. Repair plan-backed installation or wiring gaps, apply the host follow-ups,
+and retain actual failed and passing outcomes for independent review before
+changing a slot's status.
+
+The [fix list](../../evidence/artifacts/ns2604-e2e-20261004/fixes.json) ships in this
+PR. Four INTERIM token slots are held for [PR #684](https://github.com/seathatflowsinourveins/native-agent-stack/pull/684):
+context-supply, command-output, output-compression and code-index. PR #684 has not
+merged. The two open token slots have separate fixes: ccusage (FAIL, plan) and
+statusline (PARTIAL, host).
+
+The [adjudication prompt record](../../evidence/artifacts/ns2604-e2e-20261004/adjudication-prompt.md)
+contains the template as run, with run-specific placeholders. The receipt retains
+the adjudicator model, run ids, timestamps and private record hashes. Registry
+integration for all eight published files, with the hashes and mirrored fields
+this bounded repair changed, is the PR's final registry-only commit, and
+validate.py passes at that head.
+
+This repair corrects publication contradictions against adjudication.json,
+fixes.json, both workflow run records and the final 80-slot merge. The completeness
+critic covers the reviewer's/executor's 14/8 split, the four held token slots,
+the two separate open token fixes and the published prompt/fix-list handoff.
+It preserves technical strings while masking concrete home/profile paths and
+local user identities. Source inspection and structural checks supply the repair
+evidence; fresh host replication and upstream benchmark acceptance remain absent.
+
+## Alternatives and overturn condition
+
+The publication brief reports an earlier log-based readiness estimate of **62%**
+with weaker criteria. Its underlying log corpus is not supplied here, so that
+number is retained as reported comparison context. Installation or invocation
+logs can support part of a claim without proving upstream acceptance, native
+wiring and fresh client-session use. The present 38% uses the stricter retained
+reviews and final adjudications.
+
+Alternatives were to preserve that 62% log estimate as final readiness, publish
+executor labels without review, or publish only aggregate counts. The status
+chain was selected because it preserves the stricter evidence and the actual
+remaining work without exposing raw sessions.
+
+A new dated qualification would overturn the current readiness result when it
+supplies native passing acceptance, required wiring and fresh session use for
+the same 80 slots, followed by independent review and adjudication. Recovering
+original contradictory evidence could also justify a source-cited correction.
+Merging #684 or applying fixes alone does not change this historical receipt.
+A changed denominator needs an explicitly different scope and comparison.
+Provenance corrections update publication completeness without changing historical
+host readiness. Preserve this record and append later evidence.

evidence/artifacts/ns2604-e2e-20261004/adjudication-prompt.md

@@ -0,0 +1,22 @@
+# Adjudication prompt (Opus, per disputed slot)
+
+Model: `opus` (Claude Opus 5.5) at effort `max`, agent type `evidence-reviewer` (read-only), one agent per disputed slot,
+in two Claude Code workflow runs on 2026-10-04 (17 slots, then 5 more after the executor statuses were corrected from the
+workflow's schema-checked returns). Schema-forced return: slot_id, final_status, side, reason, fix, fix_kind, needs_user,
+sources. The template below is the prompt as run, with run-specific values shown as <placeholders>.
+
+```
+Adjudicate one slot of the NativeStack2604 live E2E (date 2026-10-04). Read-only; never run anything on any WSL distribution; never read credential files or values.
+
+Slot: <slot_id> (unit <unit>). The Sonnet executor proposed <executor_status>; the independent GPT Sol reviewer judged <reviewed_status>.
+
+Status definitions (from the review brief): READY = present at version, its upstream acceptance passes, wired natively, and used in a fresh headless session if the slot is client-facing. PARTIAL = some of these hold. FAIL = absent or broken. BY_DESIGN = nothing to install by design. INTERIM = an explicitly held interim the plan records (for example a token tool awaiting the token-layer PR #684); an INTERIM must not hide a missing install the current plan says should exist.
+
+Read: the executor evidence <e2e dir>/<unit>.json (this slot's record: commands, exit codes, output excerpts), the GPT review <e2e dir>/review-<unit>.json (this slot), the slot entry in <e2e dir>/units.json, and the install plan row in <host path> (origin/main).
+
+Executor summary: <executor_summary>
+Reviewer reason: <review_reason>
+Reviewer fix: <fix>
+
+Decide the status the evidence supports today, on the definitions above. Side with whichever is right, or neither. Then give the fix that would make it READY (or confirm none): exact upstream command or plan change, whether it is a repository plan change or host state, and anything that needs the user (a sign-in, an identity, a policy choice), else an empty string. Cite file:line in the evidence or plan and any upstream doc the reviewer cited.
+```