D73 + D74: OpenSandbox wire-protocol face + Intent contract

Two full cycles, delivered as a linear stack on top of 7a91a51 (v1.9.27, 23 commits).

D73 — OpenSandbox wire-protocol face (c43ea1b)

• Protocol anchored to Ali upstream main: specs/sandbox-lifecycle.yml frozen snapshot (17 endpoint families) + SDK 1.1.0 baseline
• Independent mux face (internal/schedv2/osbapi/): /v1 dual-mount + OPEN-SANDBOX-API-KEY auth + 202-async create shape (image/snapshotId/templateId)
• Tolerance matrix: gpu/resourceRequests/networkPolicy(CIDR)/secureAccess/volumes → 400 fail-closed
• execd injection trio closes the containerd-form envd isolation gap (D53-era boundary)
• Real SDK 1.1.0 direct-connect: 10 legs, 68 assertions; 3 real defects fixed (router mount / endpoint guessing / templateID carrier)

D74 — Intent contract, storage & consumption format (9396b33, v1.9.29)

• Intent is the storage AND consumption format: spec_json stores Intent JSON with kind: intent.v1 self-describing marker; read side dispatches by kind (direct / legacy SandboxSpec adapter fallback / unknown → fail-closed)
• Three invariants (now global hard constraints): I intent-persistence ⊇ user-created-intent (accept ⇒ persist); II fail-closed duality (cannot honor → 400); III single pivot
• Three adapters (e2b / osb / SandboxSpec-legacy) with per-face fidelity consistency tests
• Full downstream consumption switched: snapshot freeze, three-level resolution, agent wire, materializer, recovery chain, GC/TTL
• Field-fidelity audit (9 violations) fully resolved: 8 fixed (envVars broken since D53 — first time working; extensions/platform bucketed; entrypoint real execution — /bin/sleep 3600 hardcode abolished; labels to pod; failurePolicy; resources restore fidelity) + 1 retired (ports/DNAT — removed with native-face create, separate future cycle)
• Face-level verification: 63 steps all PASS — per-field round-trip matrix (both faces × create/restore/GET readback/execution layer): OSB entrypoint real-execution double assertion (marker + PID1 argv ≠ sleep), env verified inside container, PodAgent re-supply recovery fidelity (env + metadata + real cgroup limits), legacy fallback, snapshot-freeze buckets carry through
• 1 real defect caught by T6 and TDD-fixed (ffefe18): snapshot-consumption family dropped frozen labels/extensions

Verification

• Unit: 49 packages ok / 0 FAIL (+ new schedv2/intent); -race -count=2 on 11 touched packages all green
• Regression: TestE2BRuntime_* 34/34 PASS 0 SKIP; e2e 47 specs → 42 Passed / 5 Skipped (attempt 1)
• Cluster: rev98 = new standard state (helm values byte-identical to rev96, ledger zeroed, residue all-zero)
• v1 (scheduler/worker/limiter + agent v1 routing) zero modification (frozen baseline 06b3794)

Docs

• docs/plans/: d73-prep-wire-protocol / d73-evidence / d73-t4-evidence / d74-prep-field-audit / d74-evidence / d74-t6-evidence
• docs/manual/: 03/06 field taxonomy & custody obligations, 05 DNAT retirement notice
• Design doc v1.9.29 (external): three invariants archived

Adaptation: repo-wide rename (upstream !53)

Merged upstream/main (8a5e031, FluxSandbox → SandboxSched rename, 554 files) into this branch:

• ba6da7b — merge adaptation: 15 content conflicts resolved as semantic unions (our D73/D74 changes + renamed strings/imports); 25 new files (intent/, osbapi/, adapters) renamed-path aligned; flux-sandbox residue in Go code/manual = 0
• 12dc368 — fix(nodeagent): relay response-header whitelist made case-insensitive — a real defect introduced by the rename (Go canonicalizes X-SandboxSched-* to X-Sandboxsched-* per dash-segment; the camelCase-sensitive prefix filtered out exactly the headers it allows; caught by TestRelayForwardsHeartbeatFaithfully)

Verification ladder (post-adaptation)

• Unit: 49 packages ok / 0 FAIL; -race -count=2 on 11 touched packages all green; TestE2BRuntime 31 RUN / 0 FAIL 0 SKIP
• Cluster: redeployed with renamed chart charts/sandboxsched (release rev 99, values byte-identical to rev98; new API group CRDs *.sandboxsched.io applied; ledger five tables zeroed per e2e convention)
• e2e v2-only re-run: --ginkgo.label-filter=v2 → 4/4 PASS ×2 attempts (V2-E2E-001..004: capacity admission / HMAC authn / create-once / release), 43 v1-plane specs excluded by filter
• Honest notes: MySQL ledger data reset with renamed PVC (no evidence loss); old-API-group SSG CR manually removed; pre-existing watcher image double-name (sandboxsched/sandbox-watcher vs Makefile sandboxsched/watcher) aligned via tag