fix(engine): emit SystemMemory cgroup notice to stderr, not stdout

[jrusso1020]

What

Emit the [SystemMemory] cgroup memory limit detected: ... notice with console.warn (stderr) instead of console.info (stdout) in packages/engine/src/services/systemMemory.ts.

Why

The check --json command writes its machine-readable envelope to stdout via console.log. This SystemMemory notice was also going to stdout (console.info), so on any host with a cgroup memory limit it prepended a non-JSON line ahead of the --json payload and broke consumers that parse stdout as JSON. Standard CLI practice: stdout carries the data payload, diagnostics go to stderr. The sibling "unable to read cgroup" warning already uses console.warn (stderr); this aligns the detected-limit notice with it.

How

One-line change: console.info(...) → console.warn(...) for the detected-limit notice. Message text and all memory-adaptive behaviour are unchanged — only the stream moves from stdout to stderr. dist/ is gitignored, so there is no built artifact to update.

Test plan

Added a regression test in packages/engine/src/services/systemMemory.test.ts asserting the detected-limit notice is emitted on console.warn and NOT on console.info/console.log, so check --json stdout stays machine-parseable.

• bunx vitest run src/services/systemMemory.test.ts → 27 passed (vitest labels the notice output as stderr)

• oxfmt + oxlint on both changed files → clean

• Unit tests added/updated

• Manual testing performed

• Documentation updated (if applicable)

[vanceingalls]

RIGHT direction — approve. Root fix at the right layer: flip the SystemMemory cgroup diagnostic from stdout (console.info) to stderr (console.warn), aligning it with the sibling warnCgroupReadFailure in the same file. Every --json-emitting CLI benefits without further wiring.

Strengths

• packages/engine/src/services/systemMemory.ts:96-100 — inline comment on why (check --json payload, Video Agent parse breakage) means the invariant lives with the log site, not just in this PR body. Future edits are unlikely to reintroduce the drift.
• packages/engine/src/services/systemMemory.test.ts:255-272 — positive-and-negative pin (asserts warn called once with the notice, AND info/log NOT called). Guards against a re-drift to stdout even if someone re-adds console.info or console.log on the detected-limit branch.
• Ground-truth alignment: packages/cli/src/commands/check.ts:118-136 confirms --json writes the machine-readable envelope via console.log (stdout) and the non-JSON banner is already gated by !asJson. The bug was exactly the described stream collision.
• Ground-truth cleanliness: rg -n "console\.info" packages/engine packages/cli at head returns only the site this PR flips — no sibling console.info was left behind for a future rework cycle.

Verified premises

• Sibling warnCgroupReadFailure at systemMemory.ts:140-148 already uses console.warn. Convention exists; this PR aligns the outlier.
• No other tests depend on the notice text; no console.info spies in packages/engine. Change is fully isolated.
• CI: all required checks green at 87327381c56c31c298fe53c76cd025bb6b05cb3f (CLI smoke required, Test, Windows, Fallow, Format, Lint, Typecheck, Build). regression-shards pending — a stdout-vs-stderr stream flip on a diagnostic log doesn't touch rendering, so those shards are non-shape-related to the merge decision.

Nits (non-blocking, optional)

• nit — packages/engine/AGENTS.md (or the closest package convention doc) could carry a one-line rule: "engine diagnostics use console.warn; stdout is reserved for CLI data payloads." The inline comment gets us most of the way; a docs echo makes the convention discoverable via grep instead of by finding this file. Would prevent the next contributor from adding console.info somewhere new.
• nit — the comment mentions the Video Agent parse specifically; if the incident has a ticket / Datadog link, threading it inline would help the next debugger who greps for the failure signature. Not blocking.

Verdict: APPROVE
Reasoning: One-line stream flip at the right layer, positive+negative test pin, no sibling drift, sibling function already established the stderr convention, and the merge-gate blocker is REVIEW_REQUIRED only.

— Via