fix(test): harden injectWindow after retroactive team review

[bdchatham]

Summary

injectDeadline (renamed injectWindow here) shipped via PR #474, an unauthorized merge that never got real review. Rather than reverting it (the mechanism itself — bound fault-injection wait independently of the 40m scenario timeout — is sound), this gives it the retroactive review it never got: systems-engineer (dissenter), kubernetes-specialist, and idiomatic-reviewer, independently and blinded.

Two of three reviewers independently converged on the same real problem from different angles, plus one idiom finding:

• Latent correctness bug: the old 5-minute window exceeded the default CHAOS_DURATION (3min) — a slow-but-successful injection could return after its own fault had already self-expired, silently turning the "chain stays live under fault" check into a false green. Fixed with an enforced injectWindow < faultDur guard (fails loud) and lowered the window to 90s.
• Comment overclaim: "stuck, not slow" isn't something a blind timeout can actually establish — chaos-mesh's own retry backoff and a cold-starting daemon on a freshly scaled node can legitimately take a while without being stuck. Reworded to state this is a blast-radius bound, not a stuck-vs-slow classifier.
• Idiom: the inline comment duplicated the const's doc comment near-verbatim (and misplaced a sentence about a different function). Trimmed to match this package's established "rationale-on-const, lean-call-site" convention (oneShotObserveWindow).

One reviewer suggestion was investigated and deliberately not applied: threading the bounded injectCtx into the post-wait Get() call looked like it would close a "half-true bound" gap, but would introduce a narrow real bug instead (a fault injecting near the boundary could spuriously fail an otherwise-successful read). Kept the original ctx, documented why.

Full findings/resolution: .xreview/inject-window-retroactive-review.md.

Test plan

• gofmt -l test/integration/ clean
• go build -tags integration ./test/integration/...
• go vet -tags integration ./test/integration/...
• grep -rn "injectDeadline" test/integration/ returns nothing (full rename)
• CI (lint/test/test-integration/verify-generated)

Holding for explicit sign-off before merge, consistent with this repo's current posture on this workstream.

🤖 Generated with Claude Code

[cursor]

PR Summary

Low Risk
Integration-test harness and documentation only; no production controller or runtime behavior changes.

Overview
Hardens the integration chaos harness around fault injection timing and deferred scenarios.

injectWindow replaces injectDeadline (5m → 90s) as the cap on waiting for AllInjected, with comments reframed as a blast-radius limit rather than proof the daemon is stuck. TestChaosSuite now fails at startup if injectWindow >= CHAOS_DURATION, so a late injection cannot leave the under-fault height check running after the fault has already self-expired (false green). gateInjected keeps the post-wait status Get on the scenario ctx, not the bounded inject context, with a short note on boundary races.

disk-io-latency is removed from the active suite (IOChaos template/embed, DataDir on faultParams, and per-scenario quarantine skips). Deferred faults dns-chaos and disk-io-latency are documented in chaos_deferred_test.go so chaosScenarios stays a plain list; the disk-io rationale is updated to SeiDB path-reopen / ENOENT behavior rather than mount nesting alone.

^{Reviewed by Cursor Bugbot for commit 89a3770. Bugbot is set up for automated code reviews on this repo. Configure here.}