Update development roadmap and stabilize Windows validation

.gitattributes

@@ -76,3 +76,5 @@ provenance/compiler/predecessor/src/adaptive_inference/desktop/hermes/report_con
 # Preserve exact compiler predecessors and video qualification records.
 provenance/video-tool/** -text whitespace=cr-at-eol
 provenance/video-tool/predecessor/src/adaptive_inference/desktop/hermes/report_contract.py whitespace=cr-at-eol,-blank-at-eof
+# Exact historical CI fixture bytes include their original CRLF line endings.
+provenance/ci-fixtures/** -text whitespace=cr-at-eol

AGENTS.md

@@ -27,6 +27,7 @@ that goal. A convenient component pass does not shrink it.
 ## Shared working rules
 
 - This repository is the canonical source. Preserve engine/agent adapters as consumers of shared contracts.
+- Apply the active goal's [engineering practices](docs/multi-agent-collaboration.md#engineering-practices): small reviewable increments, explicit acceptance, meaningful validation, independent review and durable handoffs. Improve the process from evidence while prioritizing the working main workflow.
 - Keep citations, source revisions and license notices with every reused mechanism.
 - Research first before designing a solution to each new problem: search current primary research, GitHub implementations/issues/PRs and official documentation. Reuse or adapt a suitable existing mechanism before inventing one; record the sources inspected, applicability, licensing and why any necessary new work remains. Reuse prior findings when their relevance and revision are checked.
 - Use the existing queue, feedback and capability boundaries; do not invent unsupported backend controls.

CHANGELOG.md

@@ -1,5 +1,21 @@
 # Changelog
 
+## Unreleased — development after 0.4.0a1
+
+These changes belong to the development source. The published `0.4.0a1` tag and downloadable packages retain their original capabilities and validation scope.
+
+- Add **Open saved task** after a safe Stop, including tasks created later through the running controller. Accepted installed IQ3_S trials preserve the original task, confirmed reads, conversation, deadline and consumed work across application close/reopen in the same Windows boot. A completed task can be viewed without starting another worker or model, or writing its result again. See [saved-task reopening](docs/saved-task-reopening.md).
+- Add a bounded **C++17 compile-and-test recipe** that an opted-in Hermes task can request through the existing resource handoff. Reuse CMake/CTest, retained process ownership and durable tool feedback. An accepted IQ3_S foreground trial handed resources to the compiler and returned its checked result to the same task. See [the compiler recipe and evidence](docs/compiler-tool-recipe.md).
+- Add a bounded **GPU video recipe** using pinned FFmpeg/FFprobe and NVENC. An accepted IQ3_S foreground trial parked inference, encoded and checked one silent 90-frame test clip, restored the model and completed the task from saved tool feedback. This is a fixed test-clip capability; H3, audio synchronization and the full autonomous media workflow remain unqualified. See [the video recipe and evidence](docs/video-tool-recipe.md).
+- Recover missing tool feedback through the existing explicit reconciliation path when the completed result is known. Extend this to a pending journal entry when the live tool host retains the exact terminal result. Installed Hermes SDK trials use one actual video operation with scripted provider replies and resource observations; they preserve the task and deadline without replaying the operation. Unknown outcomes remain unresolved. See [completed-result recovery](docs/video-tool-recipe.md#completed-result-feedback-repair) and the [pending-effect receipt](provenance/pending-tool-recovery/installed-worker-loss-2026-11-02.json).
+- Add **automatic desktop worker replacement** for that narrowly defined pending-result boundary. In the accepted real IQ3_S trial, foreground activity parked inference, one video operation finished, and the first worker exited before recording completion. The live host reconciled the saved result and launched one replacement automatically; the same task completed within its original deadline after two total model requests. Independent closed review passed 49 checks. The separate shutdown observer failed and remains an open diagnostic; this is not active-tool interruption, whole-host loss or reboot recovery. See [automatic recovery](docs/video-tool-recipe.md#automatic-desktop-recovery) and its [physical receipt](provenance/automatic-tool-recovery/physical-foreground-recovery-2026-11-08.json).
+- Add the **durable reboot-budget foundation**: a fixed expiry and saved remaining-time cap prevent a restart from granting a new execution allowance. Source, package, installation and scripted SDK checks passed. The first actual Windows restart preserved saved work but the original allowance expired before continuation; execution was correctly refused. Successful computer-reboot continuation remains open. See [budget behavior](docs/reboot-budget.md) and the [retained expired attempt](provenance/installed-reopen/reboot-budget/physical-restart-expired-2026-11-02.json).
+- Add browser-video observation and comparison tooling, retain unsuccessful measurement attempts and correct continuation-test cleanup. The attempted comparison did not establish desktop/video performance because its observation conditions differed; no broad coexistence or speed claim follows from it.
+- Make evidence-based agent engineering an explicit goal requirement: bounded changes, meaningful acceptance, independent review, reproducible handoffs and a protected main branch. See the [collaboration practices](docs/multi-agent-collaboration.md#engineering-practices).
+- Correct two Windows CI fixture timing assumptions after retaining their original failures. Use a logical observation clock for an ordinary continuation test and a 30-second whole-fixture join guard for retained control work. Twelve targeted tests passed, including stale-observation refusals; production freshness and task deadlines are unchanged. Connect their exact retained predecessors to the existing historical-source reader after its separate hosted failure. The broader local reader sweep timed out after 118 passing checks and remains incomplete; fresh hosted validation is required. See the [fixture evidence](provenance/ci-fixtures/2026-11-09/record.json).
+
+The automatic-recovery source passed the complete **1,595-test local core suite** and all **16 additional workflow commands**, with the earlier failed test-double attempt retained. Source checks, installed SDK fixtures and physical IQ3_S evidence remain separately identified in the [evidence ledger](docs/evidence.md) and [local validation record](provenance/automatic-tool-recovery/local-validation-2026-11-08.json). These results do not complete M3, M4, M5, daily-use qualification or the full Adaptive-Hermes-Media-Autonomous Test.
+
 ## 0.4.0a1 — supported Windows reference alpha (2026-11-01)
 
 This prerelease packages the supported Windows/IQ3_S/Hermes reference, ordinary