Boot regression guards

Production HTTP and worker boots load extension metadata from bootstrap/cache/capell-package-manifests.php. Generate it during deployment with php artisan optimize or php artisan capell:package-cache. A missing or invalid cache fails a non-console boot with a remediation message; console commands retain discovery so installation and cache recovery remain available.

Admin resource-group discovery can run repeatedly during route registration, including on public requests. AdminSurfaceContributionRegistry reuses the ordered contributions for each surface type until registration, replacement or clear changes that type. Keep class and group filtering outside this cache. ExtensionContributionReceiptRegistry indexes each receipt’s seven scalar fields once; scanning and serialising every previous receipt on every insertion makes boot work quadratic in the number of contributions. Its identity index must clear with the receipts and preserve strict field equality and insertion order. The two registry unit suites cover those mutation and identity boundaries.

Three eloquent.*: * listeners are intentionally retained for created, updated, and deleted events. They support third-party Page subclasses used as error pages. ErrorPageModelInvalidationObserver immediately rejects unrelated model types, translation owners, and timestamp-only updates. The boot architecture test pins this as the complete first-party wildcard set.

Run the repeatable isolated-process benchmark with:

composer benchmark:boot -- \
    --profile=production \
    --cache=optimized \
    --iterations=25 \
    --warmups=5

--profile accepts full, production, combined, public, admin, or authoring; --cache accepts manifest, optimized, or uncached. The historical positional iteration form (composer benchmark:boot -- 10) remains supported. Use --format=json for retained evidence and --profiling to include process, framework, and per-provider register/boot timings.

combined and authoring are the immutable runtime-role profiles; production and admin remain compatibility names for existing benchmark automation. Runtime-composition admission uses a paired same-host run:

composer benchmark:runtime-roles -- \
    --cache=optimized \
    --iterations=25 \
    --warmups=5 \
    --format=json

It records combined/public p50 and p75 from the same revision and exits non-zero if public p75 regresses. See Runtime roles for the deployment and rollback contract.

Each invocation creates a disposable workspace under the operating system’s temporary directory, copies only the Testbench application skeleton, links the repository dependencies and workbench, writes its profile configuration and caches there, warms a CLI OPcache file cache there, and removes the workspace afterward. It never writes tracked workbench files. Production profiles explicitly disable Composer package discovery and omit screenshot, Pest, Ray, Pail, Octane, and other development providers.

The report includes p50, p75, p95, IQR, 10% trimmed mean, Tukey outliers, raw samples, and a fingerprint covering Git, dependencies, runtime versions, provider order, package manifests, and configuration-cache state. Warmups are not included in the reported samples. Primary samples measure in-process Laravel application creation and bootstrap so operating-system process scheduling does not distort the production-like boot result. --profiling separately reports total child-process time, process overhead, and provider register/boot timings.

Establish a baseline only after three 25-sample runs have median spread no greater than 3% and IQR/median no greater than 10%. Compare revisions with interleaved runs using the same profile, cache mode, PHP/OPcache settings, lock file, and provider fingerprint. The Git SHA and manifest hash are expected to differ when the compared source changes; all environment fields must match.

On the 2026-07-19 Phase 7.2 development machine, seven historical uncached-discovery boots measured 455.15 ms median before the original guard; seven cached boots measured 433.82 ms median afterward (21.33 ms, 4.7% lower). That figure predates these production-like profiles and remains reference-only.