docs(sync): close web end-to-end validation

AGENTS.md

@@ -17,6 +17,10 @@ Trainlog consists of:
 
 The desktop SQLite database is the canonical long-term history.
 
+The full-generation orchestrator and Web sync surface are implemented behind a
+trusted local opt-in. Default deployed automatic exchange remains V3 until the
+separate controlled-rollout decision.
+
 Android has an independent local SQLite store for offline capture. SQLite
 database files are never synchronized directly.
 

CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## Unreleased — 0.1.2
 
+- Completed the grouped `TRAINLOG_SYNC_WEB_END_TO_END_V1` mission: the opt-in
+  full-generation orchestrator, protected asynchronous Web API, global sync
+  control, durable reconnect reporting, Dashboard refresh and read-only draft
+  summary pass isolated C, Python, Android, frontend and Firefox validation.
+  Default V3, physical MTP, Drive and deployment remain unchanged.
+
 - Completed `TRAINLOG_SYNC_GENERATION_ACK_V1` with desktop schema v21 and
   Android schema v20. Added coherent per-platform capture, strict bounded
   manifest V1, immutable manifest-last publication, whole-generation SQLite

README.md

@@ -323,8 +323,9 @@ As of 2026-09-17:
   `WEB_DASHBOARD_V1=PASS/FROZEN`; the other four Web routes remain placeholders;
 - `TRAINLOG_WEB_V1=CONTRACT_FROZEN / IMPLEMENTATION_STARTED`;
 - `TRAINLOG_SYNC_GENERATION_ACK_V1=PASS/FROZEN` through explicit staged
-  services; automatic synchronization still selects V3. The next cursor is
+  services; automatic synchronization still selects V3. The opt-in grouped Web
-  `TRAINLOG_SYNC_ORCHESTRATOR_REPORT_V1`.
+  path is validated in isolation; the next boundary is
+  `TRAINLOG_SYNC_DEPLOYMENT_READINESS_V1`, not an authorized deployment.
 
 The latest executable result belongs in
 [current state](docs/current_state.md), not in multiple README narratives.

docs/architecture.md

@@ -414,6 +414,15 @@ carrier.
 
 `trainlog-syncd` is a small user-session agent.
 
+### Full-generation application orchestrator
+
+The opt-in full-generation path is owned by `sync_orchestrator.py`, launched as
+an owned process by the C Web adapter. It uses the same XDG `sync.lock` as the
+legacy engine, never borrows the Web process's SQLite handle, and accepts peer
+transport only from a trusted local configuration. The HTTP event loop only
+admits work and reads a bounded durable report. V3 remains the default when the
+trusted generation configuration is absent.
+
 It polls for a new Android request and invokes the same shared C synchronization
 engine used by the TUI.
 

docs/current_state.md

@@ -47,6 +47,10 @@ parallel implementation of its rules.
 | Web Dashboard tiles | `WEB_DASHBOARD_TILES_V1=PASS/FROZEN` |
 | Web Dashboard visualizations | `WEB_DASHBOARD_VISUALIZATIONS_V1=PASS/FROZEN` |
 | Web Dashboard V1 | `WEB_DASHBOARD_V1=PASS/FROZEN` |
+| Sync orchestrator/report V1 | `TRAINLOG_SYNC_ORCHESTRATOR_REPORT_V1=PASS/FROZEN` |
+| Web sync API V1 | `TRAINLOG_WEB_SYNC_API_V1=PASS/FROZEN` |
+| Web sync button V1 | `TRAINLOG_WEB_SYNC_BUTTON_V1=PASS/FROZEN` |
+| Sync Web end-to-end wrapper | `TRAINLOG_SYNC_WEB_END_TO_END_V1=PASS/FROZEN` (isolated) |
 | Local Web | `TRAINLOG_WEB_V1=CONTRACT_FROZEN / IMPLEMENTATION_STARTED` |
 | Complete synchronization gap contract | `TRAINLOG_SYNC_GAP_CONTRACT_V1=CONTRACT_FROZEN / IMPLEMENTATION_IN_PROGRESS` |
 | Isolated synchronization test environment | `TRAINLOG_SYNC_TEST_ENV_V1=PASS/FROZEN` |

docs/design/sync_web_end_to_end_v1_evidence.md

@@ -0,0 +1,75 @@
+# Sync Web end-to-end V1 evidence
+
+Status: `TRAINLOG_SYNC_WEB_END_TO_END_V1=PASS/FROZEN` on the development
+branch. This status means implemented and validated with isolated stores and a
+directory transport adapter. It does not mean physical MTP, Drive, deployment,
+or a merge to `main`.
+
+## Production call graph
+
+```text
+React SyncControl
+ -> POST /api/v1/sync (Host + Origin + ephemeral CSRF)
+ -> C Web admission (trusted opt-in config, one request identity)
+ -> owned sync_orchestrator.py process
+ -> common XDG sync.lock
+ -> trusted peer adapter
+ -> Android SyncGenerationService / desktop sync_generation_exchange.py
+ -> durable generation and ACK evidence
+ -> bounded run report
+ -> GET /api/v1/sync/status polling
+ -> Dashboard Core snapshot refresh after local commit
+```
+
+The deployed/default path remains V3. Full generation mode is unavailable
+unless `TRAINLOG_SYNC_GENERATION_CONFIG` names an absolute, trusted local
+configuration. HTTP input cannot select a database, executable, transport,
+capability, path, timeout, or causal bypass.
+
+## Requirement-to-proof matrix
+
+| Requirement | Production entry point | Exact proof | Observed result |
+|---|---|---|---|
+| Web admission and passive status | `web_server.c` `/api/v1/sync`, `/api/v1/sync/status` | `web_server`, `sync API` | 202/409/403/405, passive GET and CSRF header pass |
+| Async owned worker | `sync_orchestrator.py` | `sync_orchestrator`, Firefox scenario | HTTP remains responsive and child is reaped |
+| Durable typed phases | atomic `.sync-run.json` replacement | `test_sync_orchestrator.py` | correlated completion survives reopen |
+| Common contention | XDG `sync.lock` | orchestrator conflict test and legacy lock contract | second owner fails without a second import |
+| Strict compatibility | orchestrator capability set | missing-capability test | `explicitly_degraded`, missing capabilities retained |
+| Generation semantics | desktop and Android generation services | `sync_generation_exchange`, `SyncGenerationServiceTest` | coherent capture, whole transaction, ACK and replay pass |
+| Independent producers | public C fixture and Android repository fixture | `androidGenerationCapturePublishConsumeAndAckSurviveReopen` | both direction chains pass after reopen |
+| All mandatory domains | generation descriptors and apply order | generation suites and causal seven-kind test | required empty artifacts included and validated |
+| Causal immutability | `sync_causal_publications` | causal multi-generation tests | operation bytes/digests unchanged |
+| Late failure | transaction-neutral apply helpers | desktop and Android late-domain tests | prior writes roll back; no consumed ACK |
+| Lost/late ACK | durable ACK tables | generation restart tests | exact ACK reconstructed; replay unchanged |
+| UI click/reload | `SyncControl` | `test_web_sync_browser.py` | real Firefox click, completion, draft details and reload pass |
+| Draft visibility | status `drafts` projection | `SyncControl.test.tsx`, Firefox scenario | stable ID/state/type/count shown read-only |
+| Dashboard freshness | `onCommitted` callback | component test and browser flow | Core snapshot refetched after committed phase |
+| Mutation security | existing Host/Origin/CSRF plus strict body | `web_server`, `sync API` | proxy origin retained; no CORS or client paths |
+| Optional AI | separate `ai_midpoint` and `ai_post_sync` | orchestrator report validation | not-configured remains explicit, never main success |
+| Default isolation | missing trusted config | C HTTP test | button disabled and POST returns explicit conflict |
+
+The Firefox peer command in the browser scenario is the bounded directory
+transport fixture. Cross-implementation Android/desktop behavior is proven by
+the production generation suites in the same final validation run. The test
+does not claim physical-device timing, MTP object behavior, or Drive access.
+
+## Bounds
+
+- request body: existing 4 KiB HTTP limit;
+- headers: existing 8 KiB limit;
+- durable report and each worker output stream: 64 KiB;
+- trusted configuration: 16 KiB;
+- diagnostic: 1024 UTF-8 bytes;
+- draft summaries: 32;
+- peer command: 16 fixed arguments, each at most 4096 bytes;
+- trusted timeout: 1 to 900 seconds;
+- frontend polling: one request at a time, 750 ms between completions;
+- generation retention: existing eight outgoing generations per consumer.
+
+## Real limitations
+
+Android must participate through its foreground synchronization flow to create
+a publication correlated to the accepted run. No always-on Android background
+service was added. A waiting phase is therefore truthful and may require user
+action. Physical MTP and Drive remain rollout gates. A browser disconnect does
+not cancel the worker; server shutdown terminates only its owned process group.

docs/exchange_format.md

@@ -701,3 +701,13 @@ Generation membership never changes a causal V1 operation. The immutable
 operation retains null `publication_context`, its original ID and digest;
 `sync_causal_publications` records first emission and retransmission by
 operation/generation identity.
+## Operational full-generation run report
+
+The Web/application boundary uses a separate bounded operational report; it
+does not extend any frozen business artifact. It carries the stable run and
+request IDs, trigger, requested/effective mode, peer and generation IDs,
+canonical phase, monotonic progress revision, real or unavailable counters,
+per-domain coverage, bounded draft summaries, diagnostics and separate AI
+results. `completed` is accepted only after the worker supplies the correlated
+manifest/ACK evidence. The report is at most 64 KiB and is replaced atomically
+with private permissions.

docs/roadmap.md

@@ -33,7 +33,7 @@ visualizations are implemented; the other four Web routes remain placeholders.
 ## Current cursor
 
 ```text
-CURRENT_OPERATIONAL_CURSOR=TRAINLOG_SYNC_ORCHESTRATOR_REPORT_V1
+CURRENT_OPERATIONAL_CURSOR=TRAINLOG_SYNC_DEPLOYMENT_READINESS_V1
 ```
 
 The focused `TRAINLOG_SYNC_CAUSAL_DELETE_V1_CLOSEOUT` is complete before this
@@ -42,8 +42,9 @@ cross-platform draft revision identity and resource-bound admission are closed.
 `TRAINLOG_SYNC_GENERATION_ACK_V1=PASS/FROZEN`: both platforms expose explicit
 coherent capture, immutable manifest publication, whole-generation
 transactional consumption, durable correlated ACK and restart/replay entry
-points. Automatic V3 selection remains unchanged. The next cursor is only
+points. The orchestrator, Web API and Web control are now complete behind an
-`TRAINLOG_SYNC_ORCHESTRATOR_REPORT_V1`; no orchestrator work starts here.
+explicit trusted opt-in. Automatic V3 selection remains unchanged. The next
+boundary is deployment readiness; no rollout starts here.
 
 `TRAINLOG_SYNC_GAP_CONTRACT_V1=CONTRACT_FROZEN / IMPLEMENTATION_IN_PROGRESS`
 defines the complete-sync target and its bounded dependency order.

docs/sync_exchange.md

@@ -777,6 +777,18 @@ and retransmission because generation membership lives only in
 
 ## 19. Legacy active-protocol limitations
 
+The Web end-to-end tranche adds an explicit opt-in application orchestrator.
+`POST /api/v1/sync` admits one correlated run and `GET /api/v1/sync/status` is
+passive. The worker owns blocking peer/transport work, shares `sync.lock` with
+the TUI and daemon, and publishes bounded durable phases. Android foreground
+participation is required for a fresh run-correlated publication. No filesystem
+mtime or old generation can satisfy that freshness boundary.
+
+The opt-in is `TRAINLOG_SYNC_GENERATION_CONFIG`, an absolute trusted local
+configuration path. It is launcher-owned and cannot be supplied through HTTP.
+Without it, the staged path reports disabled and the deployed V3 path remains
+unchanged.
+
 The complete-sync semantics remain frozen separately in
 [`TRAINLOG_SYNC_GAP_CONTRACT_V1`](design/sync_gap_contract_v1.md). The staged
 generation slice implements its batch boundary but does not activate it.

docs/sync_rollout_readiness.md

@@ -0,0 +1,38 @@
+# Full-generation sync rollout readiness
+
+This is a non-executed rollout checklist. The development branch implements the
+path; the daily installation still uses V3.
+
+## Compatible set
+
+- desktop binary containing schema v21, the Web API and the generation worker;
+- `sync_generation_exchange.py` and all matching transaction-neutral codecs;
+- Android schema v20 with `SyncGenerationService`;
+- peers advertising manifest V1, ACK V1, causal deletion V1, history V4 and
+  execution draft V1 under their stable `peer_<uuid-v4>` identities;
+- a trusted local orchestrator configuration referencing the packaged peer/MTP
+  helper with fixed arguments.
+
+## Controlled sequence
+
+1. Back up both SQLite stores independently and record their schema versions.
+2. Stop the user services only during the separately authorized deployment.
+3. Install the mutually compatible desktop helper, binary and Android APK.
+4. Migrate synthetic copies first; migration rollback means restoring a backup,
+   not merely replacing the binary.
+5. Verify peer identities and advertised capabilities without pairing by name.
+6. Perform one non-destructive foreground Android publication and desktop run.
+7. Verify both manifest digests, both durable ACK records, lineage and readers.
+8. Keep V3 selectable until the controlled full-generation smoke is accepted.
+
+## Recovery limits
+
+Do not delete an unacknowledged generation, the last acknowledged generation,
+its previous valid generation, or causal evidence. An ACK timeout is unresolved,
+not rollback. Eight retained outgoing generations exhaust admission and require
+operator diagnosis; this tranche adds no tombstone or evidence garbage
+collection. A migrated database is recovered from its verified backup, never by
+assuming an older binary reverses its migration.
+
+Physical MTP, actual Drive configuration, service coordination, installed APK,
+main merge and deployment are intentionally not performed here.

docs/tests.md

@@ -1216,6 +1216,14 @@ does not author expected business JSON or repair artifacts.
 
 After these additions the direct normal suites report 78/78 Meson tests and
 216 Android tests: 212 passed, four historical-fixture skips, zero failures.
+
+The grouped Web end-to-end tranche raises the direct normal and sanitizer
+inventories to 79/79 and adds `sync_orchestrator`, protected sync API
+coverage, frontend sync assertions and a real headless Firefox scenario in
+`tests/test_web_sync_browser.py`. Its detailed matrix is
+[`sync_web_end_to_end_v1_evidence.md`](design/sync_web_end_to_end_v1_evidence.md).
+The browser uses the production C server and worker with an isolated directory
+transport fixture; physical MTP and Drive are not claimed.
 The publication adapter is an isolated directory/object-store double. No
 physical MTP, phone, Drive, active service, or user database participates.
 ## Causal deletion V1