fix: resolve all 14 issues from rc.17 fresh-main QA validation (#1093–#1106)

[Bojan131]

Summary

Full-surface QA of fresh main (rc.17) on a real 5-node devnet with real data surfaced 14 issues (#1093–#1106). This PR fixes all of them. Every fix was re-verified end-to-end on a clean devnet — reproduction first confirmed broken on main, then confirmed fixed on this branch. All unit suites pass (agent 1225, cli 2018, publisher 1162).

Fixes #1093, Fixes #1094, Fixes #1095, Fixes #1096, Fixes #1097, Fixes #1098, Fixes #1099, Fixes #1100, Fixes #1101, Fixes #1102, Fixes #1103, Fixes #1104, Fixes #1105, Fixes #1106

---

#1093 — P0: ACKCollector partial peer discovery permanently bricks core publishes

Repro: boot a multi-node devnet, publish from a core node shortly after startup → QuorumUnmetError / pool_below_quorum, and the node never recovers — every subsequent publish fails too.
What was wrong: the ACK candidate pool was built once from whatever peers happened to be discovered at that moment and never refreshed, so a slow-to-connect topology poisoned the pool forever.
Fix: the pool now re-resolves from knownCorePeerIds (confirmed core nodes) on each collection attempt instead of caching a partial snapshot (packages/publisher/src/ack-collector.ts, packages/agent/src/dkg-agent.ts, sync-on-connect.ts).
Verified: fresh devnet bootstrap + immediate publishes from multiple nodes — all confirmed, zero QuorumUnmetError/pool_below_quorum in any node's logs.

#1094 — P0: no working path to update a published Knowledge Asset

Repro: publish a KA, then wm/pull-from {layer:"vm"} → No VM quads found; after working around that, wm/finalize → already finalized with a different merkleRoot. The documented edit loop was completely broken.
What was wrong: (1) the VM pull query only looked in the primary data graph, but published quads live in per-KA VM graphs (…/_verifiable_memory/{author}/{number}) and context partitions; (2) pull-from never cleared the old assertion seal, so re-finalizing always collided with the stale merkleRoot.
Fix: broadened the VM lookup to scan all VM graph locations, and assertionPullFrom now deletes all seal predicates (canonical + legacy subjects) when re-opening the draft (packages/publisher/src/dkg-publisher.ts).
Verified: full edit loop on devnet — pull-from VM → write new quads → finalize → share → vm/publish returns status: confirmed as an UPDATE, and the updated content is readable from VM on publisher and replicas.

#1095 — P0: lifecycle descriptor reports contradictory state, never records published

Repro: publish a KA, fetch its lifecycle descriptor → no published event; discard an unrelated WM draft with the same name → descriptor flips to discarded while the KA is live on-chain.
Fix: the descriptor now records the published event on confirmed publish and discard no longer clobbers post-publish state (packages/agent/src/dkg-agent-lifecycle.ts).
Verified: descriptor shows published event with tx details after publish; discarding a WM draft no longer corrupts the state of the published KA.

#1096 — P1: /api/memory/search can never match Verifiable Memory

Repro: publish a KA, POST /api/memory/search scoped to VM → always 0 results even though /api/query over VM returns rows.
What was wrong: the search built graph prefixes that predate the uniform per-KA VM layout, and matched only http://schema.org while data uses https://schema.org.
Fix: corrected the VM graph prefix matching and normalized both schema.org schemes (packages/cli/src/daemon/routes/memory.ts).
Verified: memory/search over VM returns the published entities on devnet.

#1097 — P1: documented one-shot publish flow returns 500

Repro: wm/create (finalize) → vm/publish {assertionName} per SKILL.md → 500, because publish expected the data to already be in SWM while create only sealed it in WM.
Fix: publishFromFinalizedAssertion auto-promotes a sealed-but-unstaged WM draft to SWM before publishing (packages/agent/src/dkg-agent-publish.ts); SKILL.md updated to match.
Verified: one-shot create + publish by assertion name → status: confirmed on devnet.

#1098 — P1: VM divergence on replicas + rs kc-not-synced warn loop

Repro: publish on node1 with node2 subscribed → node2's VM never materializes the KA; meanwhile the Random Sampling prover floods logs with kc-not-synced retries every tick.
What was wrong: the finalization handler read SWM quads only from the bare bucket graph, but gossiped SWM copies land in per-KA graphs — merkle verification saw 0 quads and promotion silently no-opped. The RS prover retried not-yet-synced KCs with no backoff.
Fix: getSharedMemoryQuadsForRoots now uses a CONSTRUCT spanning bucket + per-KA graphs via sharedMemoryReadBothFilter (preserving literal fidelity for the merkle recompute); the prover got exponential backoff for kc-not-synced (packages/agent/src/finalization-handler.ts, packages/random-sampling/src/prover.ts).
Verified: late-joining nodes converge on the published VM data; no kc-not-synced spam in logs.

#1099 — P1: SWM clear-after-publish does not propagate

Repro: publish from SWM → publisher clears its local SWM, but every replica keeps the stale copy forever and re-serves it to late subscribers via PROTOCOL_SYNC — even re-poisoning the publisher's own SWM on resync. The update path never cleared SWM at all, even locally.
Fix: extracted clearPublishedSwmRoots (drains bucket + all per-KA SWM graphs + ownership/meta rows) and wired it into: the mint path, the update path on the publisher (dkg-agent-publish.ts), finalization replicas (finalization-handler.ts), and update replicas (update-handler.ts, including detaching roots from WorkspaceOperation rows so the sync responder stops serving them).
Verified: after publish + update, SWM row count for the roots is 0 on publisher and replicas, and a late subscriber receives no stale SWM content.

#1100 — P1: agent-profile gossip mis-decoded as publish requests (log spam)

Repro: run any 2+ node network → persistent WARN spam …does not match topic as profile gossip was fed to the publish decoder.
Fix: gossip handler dispatches by topic before decoding (packages/agent/src/gossip-publish-handler.ts).
Verified: clean devnet logs — zero mis-decode warnings across a full test session.

#1101 — P1: wm/import-file silently skips extraction for Markdown

Repro: curl -F file=@notes.md …/api/knowledge-assets/import-file → upload accepted but extraction silently skipped (curl sends application/octet-stream).
Fix: content-type detection falls back to the file extension, and skipped extractions now report an explicit skip reason (packages/cli/src/daemon/routes/knowledge-assets-import.ts).
Verified: Markdown upload reports detectedContentType: text/markdown, extraction: completed, and extracted entities are queryable.

#1102 — P2: inconsistent CG route body params (id vs contextGraphId)

Repro: context-graph/rename and /subscribe rejected the param name that sibling routes accepted.
Fix: all CG routes accept id as an alias for contextGraphId (packages/cli/src/daemon/routes/context-graph.ts); SKILL.md documents the alias.
Verified: both spellings work on rename and subscribe.

#1103 — P2: sign-join returns 200 but the join request never reaches the curator

Repro: POST …/sign-join → 200, then wait — the curator's /join-requests stays empty forever. The route is sign-only by design, but nothing said so.
Fix: the response now carries forwarded: false plus a next hint with the exact /request-join call needed to deliver the delegation; SKILL.md documents the two-step flow.
Verified: sign-join → request-join (with curatorPeerId) → request visible in curator's /join-requests → approve-join succeeds, end-to-end on devnet.

#1104 — P1: UAL duality — published UAL missing from descriptor

Repro: after vm/publish returns the minted ual, the lifecycle descriptor still only exposed reservedUal.
Fix: descriptor now surfaces publishedUal (packages/agent/src/dkg-agent-lifecycle.ts).
Verified: publishedUal present and matching the publish response.

#1105 — P1: query-remote denies everything by default

Repro: create a CG with a public on-chain accessPolicy, query it remotely from another node → denied, because the QueryHandler only consulted the local queryAccess.defaultPolicy (deny) and never the CG's own policy.
Fix: the QueryHandler resolves the target CG's on-chain access policy; public CGs are remotely queryable without local allowlist entries (packages/query/src/query-handler.ts).
Verified: remote SPARQL from node5 against node1's public CG returns the published rows (status: OK, resultCount: 1); auto-scoping guards (no explicit GRAPH clauses) still enforced.

#1106 — P2: SKILL.md drift bundle

Five doc/behavior mismatches fixed together:

1. /api/chat now accepts the intuitive peerId/message aliases for to/text (agent-chat.ts) — verified working.
2. POST /api/update docs corrected: data is inline quads, it does not read SWM; edit loop recommended.
3. WM queries without agentAddress now default to the node's primary agent wallet instead of the bare peerId namespace (dkg-agent-query.ts) — verified returning rows.
4. Documented that sub-graph names cannot contain /.
5. Documented that import-file takes contextGraphId as a multipart form field, not a query param.

---

Test plan

• Unit suites green: dkg-agent (1225 passed), dkg CLI (2018 passed), dkg-publisher (1162 passed)
• Fresh 5-node devnet bootstrap with on-chain contracts — no publish failures (bug: core nodes permanently unable to publish — partial StorageACK protocol discovery poisons the ACK peer pool (pool_below_quorum) #1093)
• Full KA lifecycle on real data: create → finalize → share → publish → pull-from VM → edit → re-finalize → update publish (bug: no working path to update a published Knowledge Asset — wm/pull-from 500s, finalize wedges, /api/update demands an attestation no API can produce #1094, bug: KA lifecycle descriptor reports contradictory state (state=discarded + status=vm-confirmed) and never records a published event #1095, bug: documented one-shot publish flow (dkg_publish / SKILL.md §4a) returns 500 — publish-by-assertionName requires an undocumented promote:true on create #1097, bug: a KA carries two irreconcilable permanent identifiers — reservedUal (chain/author/kaNumber) vs published UAL (chain/contract/tokenId) #1104)
• Cross-node convergence: replicas materialize VM, SWM drained everywhere after publish/update, late subscriber gets clean state (bug: already-subscribed core node never receives a freshly published KA (VM divergence) and its random-sampling loop spams kc-not-synced/KCNotFoundError every 5s #1098, bug: SWM clear-after-publish does not propagate — peers that catch up afterwards resurrect SWM content the publisher cleared #1099)
• memory/search over VM, remote SPARQL against public CG (bug: /api/memory/search can never match Verifiable Memory — VM graph filter still uses pre-rc.16 "_verified" prefix (+ http-only schema.org matching) #1096, bug: query-remote denies everything by default — accessPolicy:"public" CGs are not remotely queryable, and devnet ships no query-access config #1105)
• Markdown import-file extraction (bug: wm/import-file silently skips extraction for Markdown — no filename-extension fallback for content-type detection #1101)
• CG join flow end-to-end: sign-join → request-join → curator approve (bug: sign-join returns 200 but the join request never reaches the curator — no pending request, agent never added to participants #1103)
• API param aliases on chat + CG routes (bug: inconsistent CG route body params (id vs contextGraphId) and SKILL.md documents the wrong one for /api/context-graph/rename #1102, docs: SKILL.md drift — /api/chat body, /api/update body, WM query default agent, sub-graph name rules (QA pass bundle) #1106)
• Log audit: no gossip mis-decode or kc-not-synced spam (bug: agent-profile gossip mis-decoded as publish requests — full RDF payloads dumped into logs as WARN ('does not match topic "agents", ignoring') #1100, bug: already-subscribed core node never receives a freshly published KA (VM divergence) and its random-sampling loop spams kc-not-synced/KCNotFoundError every 5s #1098)

Made with Cursor

[github-actions]

Codex review skipped: filtered diff is 23982 lines (cap: 5,000). Please consider splitting this into smaller PRs for reviewability.

[branarakic]

Heads-up from #1155 (per-KA metadata trim) — these are exactly the pre-mainnet fixes that should land, and several (#1098 replica VM convergence, #1099 SWM drain) harden the same SWM/per-KA-graph surfaces #1155 restructures, so folding them together strengthens both. One cross-track interaction to coordinate, on finalization-handler.ts + the SWM materialization path:

#1155 changes two things this PR's SWM logic touches:

• The minimal per-cgId partition no longer carries dkg:status — it moved to the label _meta graph, so isAlreadyConfirmed now reads both (partition ‖ label _meta).
• KAs are collapsed to a single bare-UAL subject; sync still Merkle-verifies them via a self-map in sync-verify-worker-impl.ts + dkg-agent-utils.ts.

#1107's #1098 getSharedMemoryQuadsForRoots and #1099 clearPublishedSwmRoots assume the pre-trim partition shape. Different methods, same finalization/SWM mate — so whoever rebases second should (a) reconcile which graph carries status/SWM materialization, and (b) re-run the #1098 SWM merkle-recompute + late-joiner convergence against the collapsed bare-UAL self-map. No regression originates from the trim (every cross-reader's _meta dependency is preserved), but the SWM convergence path is worth an explicit cross-check. (Separately: #1107 and #1132 overlap each other in resolveViewGraphs WM-prefix block — independent of #1155.)