[laithalsaadoon]

What and why

Coding agents read these docs, and a page that gives two answers to one question is a
hazard. Until now the rule was the opposite: a measurement that contradicted an entry was
appended beside it, and the old one stayed. This change sets the new rule and applies it.
A measurement that contradicts an entry replaces it, the entry keeps its date, region and
API version, and git history holds the old one.

• The rule. Changed in .github/PULL_REQUEST_TEMPLATE.md (Platform claims),
  .github/ISSUE_TEMPLATE/platform_finding.md, AGENTS.md, CONTRIBUTING.md,
  site/authored/learn/operations/run-the-live-suite.md section 7, site/authored/agents.md,
  site/authored/internals/index.md, docs/STRATEGY.md, docs/README.md, and the
  drive_platform_posture docstring.
• docs/PLATFORM.md, one entry per fact. These are merged, keeping their dates: the
  duplicate WebSocket heading, the two 4096-byte runHookPayload sections, the two
  suspend/resume sections (the ticker reads as one fact: no tick lost, timestamps skip the
  suspended time), and the outside-poll measurements from 08-15 and 09-24. Also merged:
  tagging, snapshotBuild, the base-image version sections, and the two GetMicrovm
  sections, plus the suspend and resume latency from both days. Every "corrects the earlier",
  "Later evidence" and "this refresh" sentence is gone. The preamble names al2023-1 built
  from amazonlinux:2023-minimal.
• The suspended window, measured live on 2026-10-12 (us-east-1, API 2025-09-09,
  autoResumeEnabled false). suspendedDurationSeconds ends a VM that the idle policy
  suspended: TERMINATED about 60 s after suspending, then ResumeMicrovm was refused.
  Two VMs with the idle case's exact policy (maxIdleDurationSeconds 60,
  suspendedDurationSeconds 60), suspended by an explicit SuspendMicrovm, stayed
  SUSPENDED 906 s and 917 s and resumed within 20 s; so did a third with a 3600 s idle
  limit, at 1283 s. AWS describes the window for any suspended VM, so PLATFORM and
  SUSPEND-RESUME call the explicit case measured behavior AWS doesn't promise. Seven site
  pages say the same.
• Other single answers. The PT5M ready-hook timeout no longer reads as a five-minute
  platform cap (its configured timeout wasn't recorded; the model allows 3600 s).
  PROTOCOL states once when an exec has PATH and HOME. The SUSPEND-RESUME idle
  attribution is fixed (08-15 was an idle sleep, not a CPU-busy exec). Its identity text,
  and the glossary's, match the run-hook repair and the --repair-identity capabilities.
• The guard. Microvms.CorrectionTrail, an error-level Vale rule that prose:check
  runs over every maintained page, rejects correction-trail phrasing: a bold or dated
  correction, update or erratum, a re-measure next to a date, an earlier claim or
  measurement kept or superseded, "this refresh did". Its tokens match across soft line
  breaks, and ordinary words ("retained in CloudWatch", "error correction:", "the original
  description") pass. The test runs real Vale over 32 sentences, both kinds. Two passages
  it found are fixed: examples/background-coding-agent/README.md and docs/STRATEGY.md's
  "Corrections to the initial proposal", now the guidance it led to.
• Two broken heading citations fixed: .config/mise/tasks/live.toml and docs/PROTOCOL.md.

#594 merged as 944577f, and this branch is rebased onto it; it was stacked on #594. #594 rewrites PLATFORM's
capability, identity-snapshot and MMDS sections and TRUST's identity-repair and
execution-role passages, and this branch leaves them as #594 has them. The rule needs that
rewrite to pass with no exemptions, so this merges after #594, and its base then becomes
main.

Evidence

• Review fixes, rebased on feat(example): Codex SDK in a Strands Box on a MicroVM, on bedrock-runtime #594's 3c52e62 (aa5c8c8, which the rebase onto main's 944577f made e60b136 with the same tree): mise run check passed (1,799 Rust tests, 0 failed; prose 0 errors, 4,078 warnings, no rule over its ceiling). The rule's test passes 32 real-Vale cases under mise; guards:fire -- --only prose-correction-trail printed 1 of 1 fired on b5ad263, which the rebase made e680d0d with the same rule and test.
• Rebased with feat(example): Codex SDK in a Strands Box on a MicroVM, on bedrock-runtime #594 on main 78e5f1d (37af478). mise run check and mise run docs:check passed on the top of the stack built on it (the core: resume refuses an explicitly suspended VM past suspendedDurationSeconds, which the service still resumes #597 fix at c3f7bc1): 1,802 Rust tests, 0 failed; prose 4,080 warnings; 223 site tests.
• mise run check: green on a9adb04 (1,799 Rust tests, 0 failed).
• mise run docs:check: green on a9adb04 (vitest 223 passed).
• mise run prose:check: 0 errors, 4075 warnings on a9adb04 (4111 on main), no rule over
  its ceiling, and no .vale.ini exemption.
• ./tools/check-lint-coverage.py && mise exec -- ruff check . && mise exec -- ruff format --check .
  (run inside check).
• ./conformance/run_rs.py --self-test (run inside check; only a docstring changed).
• Live conformance run: not run. The suspended-window numbers come from a separate
  2026-10-12 live measurement in us-east-1, two VMs.

Guards

prose-correction-trail: changed level: error to level: suggestion in
.vale/styles/Microvms/CorrectionTrail.yml, so trail phrasing passed the gate unseen, and
test_correction_trail_phrasing_is_an_error failed. guards:fire -- --only prose-correction-trail
printed fired: prose-correction-trail on 3582b33 and again on a9adb04.

Follow-ups

• The client fix for the suspended window, core: resume refuses an explicitly suspended VM past suspendedDurationSeconds, which the service still resumes #597, owns these product lines. Each
  says a resume past the window can't work, or that the window ends any suspended VM:
  crates/microvms-cli/src/cli.rs:304, :1057, :2506; crates/microvms-cli/src/config.rs:86;
  crates/microvms-cli/src/exit.rs:257; crates/microvms-cli/README.md:51-53;
  crates/microvms-domain/src/error.rs:133; crates/microvms-app/src/sandbox.rs:35-40 and
  :1754-1760 (require_open_suspended_window); bindings/microvms-py/src/errors.rs:88-90;
  bindings/microvms-py/microvms.pyi:1383, :3850-3854; bindings/microvms-js/index.d.ts:1000,
  :2000; bindings/microvms-js/src/sandbox.rs:1148; bindings/microvms-js/src/control.rs:32;
  bindings/microvms-py/src/control.rs:42; verify/spec/core.symspec.json:844-851; and the
  generated docs/reference/cli.md:59, :170, docs/manifest.json:4626,
  docs/TRACEABILITY.md:1203-1210, docs/insights/contract-map.md:79.
• Lower .vale/warnings.json ceilings to the new counts once this and the open agentd: an image-set default user for execs and run-path starts #595 and core: resume refuses an explicitly suspended VM past suspendedDurationSeconds, which the service still resumes #597
  changes land, so each rule's ceiling is counted once.
• Binding docstrings say al2023-1 was used "from 2026-08-06 onward"; PLATFORM records it on
  2026-08-05 (bindings/microvms-js/src/sandbox.rs:256, bindings/microvms-js/index.d.ts:1651,
  bindings/microvms-py/src/sandbox.rs:386, bindings/microvms-py/microvms.pyi:310,
  conformance/harness/image.py:36).
• crates/microvms-app/src/control/microvm.rs:709 cites a PLATFORM heading that doesn't exist
  ("A MicroVM that dies during startup reports a connection error that hides the cause").
• The PT5M ready-hook reason has no recorded readyTimeoutInSeconds; one live build with a
  timeout above 300 s would settle whether five minutes is a platform cap.
• The two build-log sections ("The build log group survives Terraform", "Build logs go to")
  stay separate: they state different facts, and code cites each heading.

Platform claims

• A docs/PLATFORM.md entry changed, and it carries a date, a region, and
  an API version. A new measurement replaces any entry it contradicts,
  in this file and in other docs. Git history keeps the old one.

Scope

• This is not an orchestrator or a fork implementation
  (see docs/STRATEGY.md).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.