Clarify continuous execution of the remediation plan

docs/roadmap/implementation.md

@@ -11,12 +11,26 @@ its original audit links own defect evidence. The
 
 ## Instructions for the implementing agent
 
-Execute Stage 1 through Stage 23 in order as one sprint. Each stage is sized as
+Once the user approves this sprint, execute Stage 1 through Stage 23 in order
-one bounded Luna assignment; stages 11–12 split the cross-module diagnostic
+under that approval. Stages are bounded work units within one continuous run,
-change. Intermediate stages need not compile independently. If a check depends
+not separate prompts, approval gates, or handoffs. Proceed to the next stage
-on a later stage, record it as pending and run it when its prerequisites exist;
+without asking for confirmation after a stage, test run, or local commit.
-do not add temporary compatibility machinery to make an intermediate stage green.
+Keep progress updates concise and report the assembled result after Stage 23.
-The final source must be coherent and validated through the aligned workspace.
+Stages 11–12 split the cross-module diagnostic change for implementation clarity.
+
+Intermediate stages need not compile independently. If a check depends on a later
+stage, record it as pending, continue in order, and run it when its prerequisites
+exist. Stage completion criteria describe what must eventually be evidenced;
+they do not require stopping for dependent verification before later work can
+proceed. Do not add temporary compatibility machinery to make an intermediate
+stage green. The final source must be coherent and validated through the aligned
+workspace.
+
+Resolve ordinary implementation choices and fix failures within the approved
+scope without another design round. A real contradiction in the selected target
+or an unavoidable external blocker may require input; first investigate in-scope
+remedies and complete remaining independent work. Do not silently change the
+target, waive a relevant failing check, or claim blocked work is complete.
 
 Read [development guidance](../development.md) and the applicable
 [agent](../policy/agents.md), [architecture](../policy/architecture.md),
@@ -46,7 +60,9 @@ Record stage completion and actual verification here. Add completion evidence to
 the corresponding catalog finding only after its entire remedy is implemented
 and verified. F21 spans stages 11–12; F24 spans stages 3, 13, and 20. Preserve the
 original audits and accepted decisions. Commit completed nonempty work units with
-plain English messages under repository policy; do not publish or push.
+plain English messages under repository policy; a commit is not a user-review
+checkpoint. Related stages may form one validated commit when needed. Do not
+publish or push.
 
 ## Validation conventions
 
@@ -78,6 +94,16 @@ module files may use explicit local replacements as additional boundary evidence
 label them as local-source checks. Once appropriate releases exist, dependency
 alignment and independent validation follow the [release procedure](../release.md)
 as separate work. Publication is neither authorized nor a prerequisite here.
+Do not wait for those releases or request publication to advance or finish this
+source sprint. Record the known independent-module limitation and continue.
+
+Unavailable optional shell prerequisites or race-detector support are evidence
+limitations, not stage gates. Run the available checks, record what could not
+run, and continue. Deterministic regression checks remain required; failures in
+supported checks require investigation and in-scope correction. Revisit checks
+deferred for internal dependencies before the final report. A genuine inability
+to run required final verification must be reported as incomplete verification,
+not converted into a passed check or a request for per-stage approval.
 
 ## Stage 1 — Retain aggregate smoke cancellation
 
@@ -144,7 +170,8 @@ Findings: F03 and the reporter-lifecycle portion of F24. Source: `reporter.go`.
    presented; assert ErrReporterClosed. Check pre-cutoff accepted delivery,
    completion/terminal ordering and visibility, plus render/write failure
    evidence. Avoid sleeps and callback reentry. Run the focused reporter tests
-   with `-race -count=1` as well as ordinary focused checks.
+   with `-race -count=1` where supported, supplementing the required deterministic
+   acceptance/finalization regression. Record unavailable race support and proceed.
 6. Document the cutoff in Reporter/AtEnd GoDoc and relevant reporting guidance.
 
 Complete F03 when acceptance cannot succeed into an abandoned batch. Keep F24
@@ -570,8 +597,9 @@ Findings: all F01–F27; no new feature work.
    GOTOOLCHAIN=local GOPROXY=off go test -mod=readonly ./... ./integrations/ckconfig/... ./integrations/ckprompt/...
    ```
 
-   Run focused reporter race checks for the final reporting changes. Broaden or
+   Run focused reporter race checks where supported for the final reporting
-   repeat tests only for changed content, failures or unresolved concerns.
+   changes, recording unavailable support. Broaden or repeat tests only for
+   changed content, failures or unresolved concerns.
 3. Run independent suites from the root and each integration directory with
    `GOWORK=off GOTOOLCHAIN=local GOPROXY=off go test -mod=readonly ./...` when
    their selected dependencies support the changes. Record exact selections.