test(hedge): branch coverage >=80% on src/hedge + fix dangling rejection

[productdevbook]

Summary

Lifts branch coverage on src/hedge/index.ts from 65.71% to 91.42% (target was >=80%) AND fixes a real unhandled-rejection bug uncovered while writing the tests.

Coverage delta

Metric	Before	After
Statements	82.14%	98.24%
Branches	65.71%	91.42%
Functions	91.66%	100%
Lines	88.88%	100%

Bug fix (fix(hedge): commit)

When the external options.signal aborted during the await new Promise(...) for delayMs, the launch loop's timer-await rejected before Promise.any(winners) was reached. The in-flight primary's promise (rejected because misina sees the composed signal abort via AbortSignal.any) had no subscriber yet, so Node logged it as unhandledRejection.

The fix attaches a no-op .catch(() => {}) to each dispatched promise at creation time. The original promise still rejects normally for Promise.any downstream; the sibling handler just satisfies Node's unhandled-rejection requirement.

The regression test (external signal aborts the delay timer between firings) now asserts no unhandledRejection fires (previously it merely swallowed the dangling rejection).

Scenarios added

• Loser AbortController fires when the winner settles first.
• max cap truncates the endpoint pool to the requested count.
• External signal aborting during the delayMs wait propagates AND no longer leaks an unhandled rejection.
• delayMs > 0 with primary settling in-window hits the if (settled) break short-circuit.
• delayMs > 0 with primary still in-flight at deadline fires the backup.
• Single endpoint degrades to a plain misina call.
• init (headers) forwards to each dispatched request.
• composeOptional(undefined, internalAc) branch (no external signal).
• Pre-aborted external signal rejects the entire hedge.
• joinUrl covers all four cases: trailing-slash + leading-slash, no slashes, base-only-slash, absolute URL bypass.
• HedgeLoserError constructor exposes name and message.
• Network error on first endpoint, second wins.
• Late-winner path: a slow request that resolves after settled flips throws HedgeLoserError("not-the-winner").
• All-fail aggregate retains existing baseline test.
• Loser abort signal observable on the underlying driver request.

Unreachable branches (remaining ~8.6%)

Three defensive code paths cannot be reached without modifying src/hedge/index.ts:

1. nonLoser ?? aggregate.errors[0] fallback. Only fires when every aggregated error is a HedgeLoserError. But HedgeLoserError("not-the-winner") is only thrown from dispatchAt when settled is already true — meaning some other dispatch resolved successfully and Promise.any resolves rather than rejects.
2. primary instanceof Error false branch. All fetch/misina rejections are Error subclasses by contract.
3. composeOptional's if (!b) return a. b is always ac.signal (an internal AbortController signal), never undefined at the call site.

Test plan

• pnpm vitest run test/hedge.test.ts — 20/20 pass
• Coverage on src/hedge/** measured before/after
• pnpm vitest run — 910/910 pass
• pnpm typecheck clean
• pnpm lint clean for test/hedge.test.ts (5 pre-existing warnings elsewhere untouched)
• Bug reproducer (external abort during delayMs) confirms unhandledRejection count: 1 before fix, 0 after
• audit-fixes.test.ts > TimeoutError carries the configured timeout flake filed as Flaky test: audit-fixes 'TimeoutError carries the configured timeout' fails ~1/N runs #120 (unrelated)

[github-actions]

Bench results

⚠️ 2 misina row(s) >5% slower than base.

benchmark	base avg	head avg	Δ
createMisina cold start :: createMisina()	484.2 ns	596.3 ns	🔴 +23.2%
createMisina cold start :: createMisina({ use: [bearer] })	536.0 ns	573.1 ns	🔴 +6.9%
createMisina cold start :: createMisina({ use: [bearer, tracing, breaker] })	1.83 µs	1.87 µs	· +2.5%
steady-state GET (200 OK, JSON parse) :: misina	296.04 µs	302.28 µs	· +2.1%
POST JSON body :: native fetch	357.28 µs	363.09 µs	· +1.6%
steady-state GET (200 OK, JSON parse) :: axios	310.12 µs	308.74 µs	· -0.4%
steady-state GET (200 OK, JSON parse) :: native fetch	324.39 µs	319.60 µs	· -1.5%
steady-state GET (200 OK, JSON parse) :: ofetch	266.75 µs	261.07 µs	· -2.1%
POST JSON body :: axios	322.81 µs	313.63 µs	· -2.8%
plugin overhead (no plugins / 1 hook plugin / 3 plugins inc. wrapping) :: misina — bearer + tracing + breaker	305.20 µs	295.46 µs	· -3.2%
POST JSON body :: misina	523.39 µs	500.97 µs	· -4.3%
hooks overhead (no hooks vs 5 hooks) :: misina — no hooks	291.91 µs	278.26 µs	· -4.7%
steady-state GET (200 OK, JSON parse) :: ky	299.00 µs	284.36 µs	· -4.9%
plugin overhead (no plugins / 1 hook plugin / 3 plugins inc. wrapping) :: misina — no plugins	281.20 µs	266.15 µs	🟢 -5.4%
hooks overhead (no hooks vs 5 hooks) :: misina — 5 hooks	294.29 µs	277.92 µs	🟢 -5.6%
plugin overhead (no plugins / 1 hook plugin / 3 plugins inc. wrapping) :: misina — bearer	282.64 µs	266.27 µs	🟢 -5.8%
POST JSON body :: ofetch	353.06 µs	332.39 µs	🟢 -5.9%
POST JSON body :: ky	480.01 µs	451.25 µs	🟢 -6.0%
retry on 503 → 200 :: misina — retry 1× then 200	1.84 ms	1.72 ms	🟢 -6.6%

_{Local Node HTTP server, Apple-class GitHub runner. Numbers fluctuate ±2-3% from runner heat alone — only sustained >5% deltas are signal.}

[productdevbook]

LGTM. Coverage lifted to branches 91.42% / lines 100% / funcs 100%. Spot-checked test assertions: mocks are tight (specific URLs, specific status codes, response bodies), assertions are specific (exact seen sequences, exact host strings, instanceof checks), no loose expect.anything(). The flagged dangling-rejection bug was real and reproducible (Node logged unhandledRejection) — fixed in a separate fix(hedge): commit by attaching a sibling .catch(()=>{}) to dispatched promises. Regression test now asserts no unhandled rejection. Pre-existing audit-fixes flake filed as #120.