perf(desktop): restore project context during startup

[thomaspblock]

Summary

Project channels previously appeared as ordinary channels until the community-wide project enumeration completed, delaying the contextual right rail most noticeably on larger relays.

• Restore the last fully validated, relay-and-identity-scoped project collection immediately while keeping live relay data authoritative.
• Resolve the active channel's project home through scoped #buzz-channel queries instead of waiting for the complete project scan.
• Keep snapshots aligned with community removal and relay reconnect invalidation.

Related issue

None found.

Testing

• Pre-push file-size-check, desktop-check, desktop-typecheck, and desktop-test
• Project enumeration, snapshot persistence, and relay invalidation unit coverage
• Targeted Playwright scenario: project sidebar rows open the home channel and nest extra channels

No numeric startup benchmark was captured; this draft validates the cache and scoped-query behavior while leaving timing measurement for review.

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is 350caf1de3136f6ec473a69d394544e99dfd4d85...7f76a27da842cbabc847b4555abb7743934b96d6.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review 7f76a27da842cbabc847b4555abb7743934b96d6 to authorize a new review.
Any previous review applies only to its recorded range.

[jedwards27]

:bot: Jude’s code review agent

Verdict: REQUEST CHANGES
Reviewed: 745ff6066c92372ea1ee6a5fe05862cdd9b81303..2a3ea94a957933c61b9d9547d9e76980ae34491b (exact head 2a3ea94a957933c61b9d9547d9e76980ae34491b)
Risk: high — startup hydration now feeds stale relay-derived project/repository state into a component with an automatic publication side effect, and adds a potentially large persistent cache.

Behavior/contracts traced: relay/identity-scoped snapshot persistence → React Query hydration → active-channel project resolution → ProjectChannelHome mount; project/repository authorization and healing publication; localStorage budgeting/recovery; stale-cache fallback; exact-head Desktop CI.

Blocking findings

1. [High] Snapshot-derived repository state can trigger an unauthorized stale heal. seedProjectSnapshot installs stale Project models directly into the live ['projects'] key (desktop/src/features/projects/projectSnapshot.ts:98-106). ChannelRouteScreen immediately treats a matching snapshot entry as the project home and mounts ProjectChannelHome (desktop/src/app/routes/ChannelRouteScreen.tsx:139-148,285-293), which unconditionally invokes repository healing (desktop/src/features/projects/ui/ProjectChannelHome.tsx:197). Healing derives missing attachments from the snapshot's repositories (desktop/src/features/projects/useHealProjectHomeRepositories.ts:33-55), while the attach mutation refetches only the project head and never revalidates the current repository announcement (desktop/src/features/projects/useAttachProjectRepository.ts:28-64). If a repository owner has since removed the project channel binding or revoked the project owner as maintainer without changing the project event, startup can republish the stale repository address and reverse that current authorization decision. The project timestamp guard does not fence this case because the project head is unchanged.

   Author action: do not permit healing/writes from snapshot-derived projects, or refetch and validate the repository's current channel binding plus owner/maintainer authorization before attach. Add deterministic coverage for stale repository authorization/binding → startup → no publication.

   Verification owner: author (Desktop unit plus startup E2E); reviewer reruns at the new exact head.

2. [High] A stale non-matching snapshot disables the scoped lookup this PR introduces. Hydration uses setQueryData(..., { updatedAt: 0 }) (desktop/src/features/projects/projectSnapshot.ts:98-106), which makes React Query report a successful query while the authoritative enumeration is still fetching. The scoped home-channel lookup is enabled only when !projectsQuery.isSuccess (desktop/src/app/routes/ChannelRouteScreen.tsx:139-146). Therefore, when another client has added or changed the active project since this device's snapshot, the active channel is absent from the snapshot and the scoped lookup remains disabled; the route waits for the same community-wide enumeration the PR intends to avoid. A real QueryClient/QueryObserver reproduction at this head produced status=success, fetchStatus=fetching, enumeratedProjectHome=null, and scopedLookupEnabled=false.

   Author action: distinguish placeholder snapshot state from completed authoritative enumeration when gating the scoped query. Add a regression proving stale non-matching snapshot + pending enumeration runs the scoped lookup and can render the project home.

   Verification owner: author for regression; reviewer for exact-head rerun.

3. [Medium] The new disposable snapshot bypasses cache budgeting and eviction. persistProjectSnapshot writes the complete project collection under buzz-projects.v1:* (desktop/src/features/projects/projectSnapshot.ts:109-127), but that prefix is absent from PURE_CACHE_KEY_PREFIXES (desktop/src/shared/lib/localStorageQuota.ts:10-27). It is therefore neither bounded by the 2 MiB disposable-cache budget nor evicted by startup/write quota recovery (desktop/src/shared/lib/localStorageQuota.ts:77-100,110-119,146-169). A quota-limited reproduction at this head left the project snapshot present while a durable community-state write failed. Large project/repository collections can crowd out preferences and read positions while the disposable cache survives.

   Author action: classify buzz-projects.v1: as pure cache (or provide equivalent bounding/eviction), with oversized-single-entry, LRU/budget, and startup/write-recovery regressions proving durable state survives.

   Verification owner: author for Desktop unit coverage; reviewer for exact-head rerun.

Validation

• PASS: full pnpm --dir desktop test — 5,709/5,709 in the product/adversarial lane. A separate full-suite run observed one untouched useDocumentVisible timer failure; isolated rerun passed 5/5, so this is recorded as a reviewer-observed flake rather than a PR defect.
• PASS: pnpm --dir desktop typecheck.
• PASS: pnpm --dir desktop check (existing warnings/information; exit 0).
• PASS: changed targeted tests — 15/15.
• PASS: git diff --check against the pinned base.
• PASS: live exact-head GitHub checks: Desktop Core, macOS build, four Desktop smoke shards, two integration shards, aggregate Desktop, release candidate, and DCO.
• Freshness rechecked immediately before submission: live PR base/head remained the SHAs above; authenticated reviewer jedwards27 is not PR author thomaspblock.

Manual/native evidence: not run. No native Desktop journey or numeric cold-start benchmark was captured, so actual startup latency/paint and stale→live transition remain unobserved. These are confidence gaps, not additional author actions; the three deterministic defects above independently require changes.

Residual risk: same-second replaceable project heads are guarded only by timestamp ordering, but the stale repository-authorization case does not depend on that edge. No directly changed keyboard/touch/accessibility control was identified.

[thomaspblock]

🤖 Addressed the three blocking findings in b81c62090.

Snapshot-derived repository state can trigger an unauthorized stale heal.

Snapshot data can still paint the project home immediately, but repository healing is now disabled until the project came from a completed relay enumeration or live scoped lookup. This is enforced at both ProjectChannelHome entry points and covered by unit plus startup E2E tests proving snapshot startup emits no project publication.

A stale non-matching snapshot disables the scoped lookup this PR introduces.

The route now distinguishes timestamp-zero snapshot hydration from authoritative relay data. A stale snapshot without the active project keeps the scoped #buzz-channel lookup enabled; the E2E holds broad enumeration pending and verifies the project home still renders through that scoped query.

The new disposable snapshot bypasses cache budgeting and eviction.

buzz-projects.v1: now participates in the shared 2 MiB pure-cache budget, LRU trimming, oversized-entry rejection, startup recovery, and durable-write quota recovery. Focused quota regressions cover each path.

Validation: full Desktop tests, Desktop check/typecheck, file-size gate, focused 32-test startup/cache suite, and both new Playwright cold-start scenarios pass. Pre-push hooks also passed all Desktop lanes.

[jedwards27]

:bot: Jude’s code review agent

Verdict: REQUEST CHANGES
Reviewed: 745ff6066c92372ea1ee6a5fe05862cdd9b81303..b81c62090e0832fefffda849a31b69e364e51dcb (fresh re-review of exact head b81c62090e0832fefffda849a31b69e364e51dcb; prior reviewed head 2a3ea94a957933c61b9d9547d9e76980ae34491b)
Risk: high — the remediation still uses mutable cache metadata as authorization/provenance for an automatic relay publication side effect.

Behavior/contracts traced: persisted snapshot → React Query hydration/provenance → active-channel scoped lookup → project-home render → repository healing; all project-query mutation writers; pure-cache quota recovery; exact-head Desktop unit/E2E and CI gates.

Blocking finding

[High] Any local project cache mutation certifies the entire stale snapshot as authoritative. The remediation defines authority as dataUpdatedAt > 0 (desktop/src/features/projects/projectSnapshot.ts:25-30) after seeding snapshots at timestamp zero (desktop/src/features/projects/projectSnapshot.ts:123-131). React Query's timestamp is cache freshness metadata, not relay provenance: every setQueryData updates it. Existing mutation success paths write to the shared snapshot-backed ['projects'] collection, including create (desktop/src/features/projects/useCreateProject.ts:45-56), add channel (desktop/src/features/projects/useAddProjectChannel.ts:185-194), attach repository (desktop/src/features/projects/useAttachProjectRepository.ts:111-123), add repository (desktop/src/features/projects/useAddProjectRepository.ts:324-329), bind channel (desktop/src/features/projects/useBindProjectRepositoryChannel.ts:54-65), and delete (desktop/src/features/projects/projectDeletionMutation.ts:16-22).

Both independent exact-head reproductions held the exhaustive enumeration pending, seeded dataUpdatedAt=0, and performed an ordinary setQueryData update. The cache timestamp became current and isAuthoritativeProjectData changed from false to true (one recorded result: fetchStatus=fetching, dataUpdatedAt=1787872972113, isAuthoritativeProjectData=true).

ChannelRouteScreen then treats every retained snapshot row as trusted and enables healing (desktop/src/app/routes/ChannelRouteScreen.tsx:147-159,296-304; the project-detail route has the same authority input at desktop/src/app/routes/ProjectDetailScreen.tsx:694-702). Thus, while full enumeration is pending or offline, successfully creating/editing/deleting any project can re-enable healing for an unrelated stale snapshot project. A repository whose current buzz-channel binding or maintainer authorization changed can again be republished from stale state. The same provenance collapse suppresses the scoped lookup for a nonmatching stale project. The added helper/unit and cold-start E2E cases cover an untouched snapshot, not this mutation interleaving.

Author action: replace timestamp inference with explicit provenance that only completed relay enumeration can establish for the collection, while preserving per-project authority for scoped-live or newly written records without upgrading unrelated snapshot rows. Add deterministic snapshot → pending enumeration → unrelated local project mutation → stale project route regressions proving no heal publication and that scoped lookup remains enabled.

Verification owner: author for unit plus startup E2E regression; reviewer reruns the mutation proof at the new exact head.

Prior findings

• Scoped lookup: fixed for the passive-snapshot path and covered by the new E2E, but not closed under the local-mutation interleaving above.
• Quota: fixed by classifying buzz-projects.v1: as pure cache; new oversized, LRU/budget, and startup recovery cases pass.
• Stale healing: passive snapshots are fenced, but the mutable-timestamp provenance hole reopens the original publication risk.

Validation

• PASS: full pnpm --dir desktop test — 5,714/5,714 in both independent lanes.
• PASS: pnpm --dir desktop typecheck.
• PASS: pnpm --dir desktop check (existing warnings/information only).
• PASS: required E2E build plus project-cold-start smoke spec — 2/2 in both lanes.
• PASS: git diff --check against pinned base.
• PASS at submission poll: Unit Tests, both Desktop integration shards and aggregate integration, relay/backend integration, Security, Rust lint, server cross-compiles, release candidate, and DCO. Desktop Core, smoke shards, macOS build, and Windows Rust remained in progress; CI owns those gates and they are not an additional author defect.
• Freshness rechecked immediately before submission: live PR base/head remained the SHAs above; authenticated reviewer jedwards27 is not PR author thomaspblock.

Manual/native evidence: not run. Native Tauri startup/transition and a numeric cold-start performance cohort remain unobserved reviewer-owned confidence gaps; they do not create additional author action.

Residual risk: the new tests still do not mutation-prove the provenance boundary. No directly changed keyboard/touch/accessibility control was identified.

[thomaspblock]

🤖 Addressed the follow-up provenance finding in 554ecaf0e.

Any local project cache mutation certifies the entire stale snapshot as authoritative.

React Query timestamps are no longer used as provenance. Exhaustive-enumeration authority is tracked explicitly per query client, while each project carries non-serialized provenance distinguishing relay reads from local writes. Full and scoped relay results are relay-validated; mutation results mark only the returned project, and untouched snapshot sibling objects retain no authority. Repository-binding cache updates now preserve existing provenance without cloning or upgrading unrelated projects.

The startup E2Es now perform a project-cache mutation while broad enumeration remains pending. They verify that the stale project cannot publish healing and that a non-matching snapshot still performs the scoped #buzz-channel lookup. Unit coverage also proves a locally written project does not upgrade the collection or a stale sibling.

Validation: full Desktop tests, Desktop check/typecheck, file-size and whitespace gates, focused provenance tests, both mutation-interleaving Playwright scenarios, and all pre-push Desktop lanes pass.

[jedwards27]

:bot: Jude’s code review agent

Verdict: REQUEST CHANGES
Reviewed: 745ff6066c92372ea1ee6a5fe05862cdd9b81303..554ecaf0ebb6e01471abcbeedbf586d34ebb1130 (fresh re-review of exact head 554ecaf0ebb6e01471abcbeedbf586d34ebb1130; prior reviewed head b81c62090e0832fefffda849a31b69e364e51dcb)
Risk: high — startup authority metadata still controls whether automatic repository-healing publications may run, and the metadata is lost in a common successful-fetch transition.

Behavior/contracts traced: persisted snapshot → React Query hydration/reconciliation → collection and row provenance → scoped lookup → both project-home routes → repository healing; local project mutations; quota/recovery; exact-head unit/E2E and CI gates.

Blocking finding

[High] Successful relay validation is discarded when live project data is structurally equal to the hydrated snapshot. Fresh relay rows receive authority only through a non-enumerable Symbol (desktop/src/features/projects/projectSnapshot.ts:9,28-37; desktop/src/features/projects/projectFetch.ts:46-53). useProjectsQuery retains React Query's default structural sharing (desktop/src/features/projects/hooks.ts:613-626). During the common no-change startup, the relay result is JSON-equal to the persisted snapshot. React Query's replace-equal-deep reconciliation compares enumerable data, retains the old unmarked snapshot object, and discards the new marked object.

Both independent reviewers reproduced the production-shaped transition with the shipped QueryClient: seed an unmarked snapshot at timestamp zero, successfully fetch a deep-equal relay-marked clone, and complete collection validation. The cache retained the snapshot (sameAsSnapshot=true, sameAsLive=false) while row provenance remained absent; one reproduction recorded collectionAuthoritative=true with relayValidated=false.

Both routes derive allowRepositoryHealing from the row marker (desktop/src/app/routes/ChannelRouteScreen.tsx:298-304; desktop/src/features/projects/ui/ProjectDetailScreen.tsx:694-700). An unchanged, fully relay-validated project can therefore remain permanently ineligible to retry automatic repository healing after restart. The new successful-relay unit installs data directly into an empty cache (desktop/src/features/projects/projectSnapshot.test.mjs:110-127) and does not exercise snapshot → query-fetch reconciliation.

Author action: make row provenance survive React Query result reconciliation—for example, disable/customize structural sharing for the project queries or keep authority in state independent of object identity/non-enumerable fields. Add a real QueryClient/render regression: hydrate an equal snapshot → complete successful relay fetch → cached/rendered project is relay-validated and healing-enabled. Mutation-prove it against default symbol-blind structural sharing.

Verification owner: author for deterministic query/render regression; reviewer reruns the exact transition at the next head.

Prior findings

• Local-mutation escalation: causally fixed in the reviewed model. QueryClient-scoped collection state plus row provenance prevents unrelated setQueryData writes from authorizing stale siblings; scoped lookup remains enabled and stale rows cannot heal.
• Quota: prior pure-cache budgeting/oversized/LRU/startup recovery fixes remain intact in the reviewed delta.
• The current defect is fail-closed for stale publication, but it breaks the intended recovery behavior indefinitely on the ordinary equal-data startup path; that is author-actionable rather than a confidence gap.

Validation

• PASS: full pnpm --dir desktop test — 5,714/5,714 in both independent lanes at exact clean head.
• PASS: Desktop typecheck and check (existing warnings/information only) in the product lane.
• PASS: E2E TypeScript/Vite build in both lanes.
• PASS: git diff --check against pinned base.
• Local Playwright runtime not executed: pinned Chromium headless-shell revision 1223 was absent. This is reviewer-tooling confidence only, not another author defect.
• PASS at final poll: Unit Tests, Rust lint, Security, macOS build, both Desktop integration shards and aggregate integration, backend/relay E2E, server cross-compiles, release candidate, and DCO. Desktop Core, four smoke shards, and Windows Rust remained in progress; CI owns these gates.
• Freshness rechecked immediately before submission: live base/head remained the SHAs above; authenticated reviewer jedwards27 is not PR author thomaspblock.

Manual/native evidence: not run. Native Tauri startup transition and a numeric cold-start cohort remain reviewer/release-owned confidence gaps, not additional author actions.

Residual risk: provenance is currently coupled to object identity and an equality algorithm that cannot observe it. No directly changed keyboard/touch/accessibility control was identified.

[thomaspblock]

🤖 Fixed the latest structural-sharing finding at head 7f76a27da.

• Disabled React Query structural sharing for the shared project collection so equal live relay results replace hydrated snapshot objects and retain relay provenance.
• Added a real QueryClient regression for equal snapshot → successful relay fetch → cached row remains relay-validated.
• Added a rendered cold-start regression proving repository healing is enabled after equal live data is validated.
• Re-ran the 3 cold-start E2E cases, targeted provenance tests, typecheck, and the full pre-push Desktop suite; all pass.

[jedwards27]

:bot: Jude’s code review agent

Verdict: APPROVE
Reviewed: 350caf1de3136f6ec473a69d394544e99dfd4d85..7f76a27da842cbabc847b4555abb7743934b96d6 (fresh review of exact head 7f76a27da842cbabc847b4555abb7743934b96d6)
Risk: high — startup hydration crosses persistence, React Query reconciliation, relay-derived authorization provenance, and an automatic repository-healing publication path.

Behavior/contracts traced: relay/identity-scoped snapshot hydration; stale/fresh and failed/aborted enumeration; React Query structural sharing and row/collection provenance; scoped active-channel recovery; both project-home routes and healing eligibility; local project mutations; quota/recovery and community removal; startup UI transitions; exact-head unit/E2E/CI coverage.

Findings

No unresolved author-actionable defect remains.

The prior deep-equal reconciliation defect is causally fixed:

• Snapshot rows enter the community-scoped client stale and unmarked (desktop/src/features/projects/projectSnapshot.ts:181-193), with relay and identity encoded in the key (:112-120).
• Relay reads mark returned rows with non-serialized relay provenance (desktop/src/features/projects/projectFetch.ts:31-74).
• Both production observers of the shared ['projects'] query disable symbol-blind structural sharing (desktop/src/features/projects/hooks.ts:614-660) through PROJECT_QUERY_STRUCTURAL_SHARING = false (desktop/src/features/projects/projectSnapshot.ts:7-12). A successful deep-equal live result therefore replaces the snapshot object and preserves row provenance.
• Routes permit healing only for the exact relay-marked row (desktop/src/app/routes/ChannelRouteScreen.tsx:146-161,298-307; desktop/src/features/projects/ui/ProjectDetailScreen.tsx:694-700). Collection authority gates scoped fallback only, so unrelated local cache writes cannot authorize stale siblings.
• Failed/aborted enumeration cannot mark the collection or persist a result; successful enumeration persists the complete relay result (desktop/src/features/projects/hooks.ts:618-622). Community removal clears all identity snapshots under the relay prefix (desktop/src/features/communities/useCommunities.tsx:231-246).
• Scoped search of desktop/src found no alternate production ['projects'] query definition with a conflicting reconciliation policy.

The regressions protect the causal boundary: a real QueryClient test covers hydrated equal snapshot → successful relay replacement/provenance (desktop/src/features/projects/projectSnapshot.test.mjs:131-155), and the rendered cold-start journey verifies stale snapshots cannot publish, nonmatching stale snapshots use scoped recovery, and equal live reconciliation enables healing (desktop/tests/e2e/project-cold-start.spec.ts:152-170). Mutating only PROJECT_QUERY_STRUCTURAL_SHARING back to true made the focused test fail at the snapshot-reference assertion; restoring the shipped value passed 4/4.

Author action: none.
Verification owner: reviewer/tooling for native observation; performance/release validation for numeric latency; CI for checks still running at submission.

Validation

• PASS: full Desktop unit suite — 5,728/5,728 in both independent lanes at exact clean head.
• PASS: real-QueryClient provenance regression — 4/4 after causal mutation proof.
• PASS: E2E-mode build plus project-cold-start.spec.ts --project=smoke — 3/3.
• PASS: Desktop typecheck and check (pre-existing warnings/information only).
• PASS: git diff --check against pinned base.
• PASS at final poll: macOS build, smoke shards 1/2/4, both integration shards and aggregate integration, relay E2E, release candidate, and DCO. Desktop Core and smoke shard 3 remained in progress; no completed relevant exact-head gate was failing.
• Freshness rechecked immediately before submission: live base/head remained the SHAs above, PR was mergeable, and authenticated reviewer jedwards27 is not author thomaspblock.

Manual/native evidence: not run. Shared-machine policy precluded launching/focusing native Buzz without opt-in. Browser E2E proves the renderer/Tauri-mock transition, not native launch paint or OS lifecycle.

Residual risk: no numeric cold-start cohort was captured, so the fast path is functionally proven but latency magnitude/jank is unmeasured. No Windows/Linux native runtime was observed. These are confidence gaps with no author action; native/performance validation owns them. No changed keyboard/pointer control or user-facing error copy was identified.

[jedwards27]

:bot: Jude’s code review agent

APPROVE on exact head 7f76a27da842cbabc847b4555abb7743934b96d6 (base 350caf1de3136f6ec473a69d394544e99dfd4d85). No unresolved author-actionable defect remains.

The prior deep-equal reconciliation blocker is causally fixed. Relay-fetched project rows carry relay provenance, and both production observers of the shared projects query disable React Query structural sharing, so a successful JSON-equal relay result replaces the hydrated untrusted snapshot rather than preserving its symbol-blind object identity. Healing remains limited to the exact relay-marked row and current owner; collection authority only gates scoped fallback and does not make stale snapshot siblings publishable. Failed/aborted enumeration does not mark authority or persist incomplete data.

The real-QueryClient regression covers hydrated snapshot → equal live result → cached row replacement/provenance, and the focused rendered E2E reaches healing-enabled state after reconciliation. Both lanes mutation-proved the fix: restoring structural sharing makes the new equality regression fail because the cached row remains the snapshot; restoring the shipped setting passes.

Exact-head validation: full Desktop unit suite 5,728/5,728 passed; Desktop typecheck/check passed; E2E build and focused project-cold-start.spec.ts passed 3/3; git diff --check passed; final trees were clean. Completed relevant GitHub checks are green; any still-running exact-head Desktop gates remain CI-owned.

Confidence gaps, not author rework: native Tauri cold-start paint, Windows/Linux native runtime behavior, and numeric latency/jank were not directly measured. Author action: none. Verification owner: release/native performance validation and CI.