Verdict: approve-worthy — self-authored, so COMMENT (GitHub blocks self-approve). 0 blockers · 0 majors · 1 minor (hardening) · confirmations below. All CI green (check, Cloudflare Pages, town-crier/gate). ADR-0037 implemented faithfully.

Confirmed correct (the load-bearing points)

• Reference identity holds. http.ts stores const middleware = opts?.guard === false ? fn : guarded(fn, onMiddlewareError) and passes that same reference to both push() and unregister(). This is the trap the naive version falls into (push guarded(fn), unregister fn → leak); it's handled correctly on all three registrars.
• Default-on semantics are strict. opts?.guard === false ? fn : guarded(...) — only an explicit false opts out; undefined / {} / {guard:true} all guard. Correct.
• onMiddlewareError wiring — service-level handler captured once, threaded into every guarded() call, undefined ⇒ defaultOnError. Correct.
• Peer cascade is additive — every sibling widens with || ^0.6.0 (never drops a prior range) + patch-bumps. fs-form #148 is already merged (base = fs-form 0.1.0 / fs-http 0.5.0), so the 0.1.0→0.1.1 bump here is clean with no merge-order dependency.

Minor — guarded()'s own error handler is unguarded (hardening)

guarded.ts catches fn(arg) but calls onError(error) outside the try/catch:

return (arg: T) => {
    try { fn(arg); }
    catch (error) { onError(error); }   // ← if onError throws, it escapes the wrapper
};

The default defaultOnError (console.error) can't throw, so the default path is airtight. But this PR makes guarded() the fleet-wide default and documents onMiddlewareError as the hook to "route to an error tracker (Sentry, kendo-error-tracker)" — i.e. it actively steers consumers toward exactly the handlers most likely to throw (network, serialization, a null-deref in the reporter). A throwing onMiddlewareError re-opens the precise interceptor-chain corruption ADR-0037 exists to close, one level up.

The doc contract says "Must not re-throw" — but that's a documented obligation, not an enforced one, on a primitive whose entire value proposition is "never lets a throw corrupt the chain." Belt-and-suspenders: wrap the onError(error) call too (fall back to defaultOnError, or a silent no-op after a last-ditch console.error). Pre-existing code (shipped in 0.5.0), but this PR is what elevates it to the default path, so it's fair to raise here rather than defer. Non-blocking — the default is safe and the contract is documented; your call whether to harden now or leave it as a doc obligation.

Notes (no action)

• Guard-by-default is technically a behavior change (a middleware throw that previously propagated is now loud-swallowed) shipped as a pre-1.0 minor — correct call: it's backward-safe direction (a no-throw middleware sees zero change) and ADR-ratified, and pre-1.0 minor matches the peer-cap discipline.
• Redundant double-wraps on WR-0290 / WR-0078 consumer sites after they bump — confirmed harmless (inner catch fires, outer never sees a throw); deferred cleanup is the right disposition.

Automated war-room agent review — posted because this PR carries the Agent Review Requested label.

Auto-approved — Town Crier verdict PASS @Head, CI green, no open MAJOR+ thread. Our approval is our independent vote (approve-alongside): a peer's review / CHANGES_REQUESTED never withholds it — we verify every blocker ourselves, and a real one drops our own verdict below PASS. Any open 🟡 MINOR threads alongside this approval are non-blocking (TC-0036 R4) — author's choice: fold here or follow-up; if folding into this PR, say so in-thread and disarm auto-merge. See the verdict comment + inline notes.