frontend-patterns/research/reviews/NavigationHistoryExperiment.md

First prospective navigation cycle — 2026-09-27

Author/reviewer: Codex, author-model; no independent human validation. Stream B, LEARN-0011. Registered protocol committed as f3697c1 before implementation/execution. Runner and reproduction, complete receipt, partial first run.

Evidence and observations

Real htmx 2.0.11 in an authored public-catalog server; Playwright 1.63.0, Chromium 151.0.7922.34 and Firefox 155.0. Dependencies locked, inputs and browser binaries hashed in the complete receipt. Existing htmx history/cache documentation was consulted for mechanism context (accessed local date 2026-09-27, living revision unknown); execution is evidence only for the pinned package and declared server policy.

Claim	Registered checks	Observation in both browsers
NH01	N01/N02	Direct filtered entry, native link/GET form, enhanced navigation, reload and fresh direct entry preserve selection; document entries include the shell
NH02	N03/N04	Snapshot back/forward makes no additional article request; evicting the pinned sessionStorage history cache causes a history-restore request and full response
NH03	N05/N06	Enhanced 503 leaves the old committed URL/content; explicit native recovery works. A failed htmx script load leaves native link/form usable
NH04	N07/N08	Correct Vary selection returns full versus fragment content in both population orders; N08's final full fetch recontacts the server, not proof that both variants stay cached
NH05	N09	Without Vary, the browser reuses the cached fragment as a full document, loses the shell, and sends no additional request
NH06	N10	Explicit Node requests receive 304 for a matching representation validator and 200/full body for the other representation; a history-restore selector returns full content

20/20 registered rows met expectations, including the intentionally broken N09. There are 18 browser cases and the same explicit HTTP sequence repeated twice; do not describe N10 as browser automatic revalidation. Snapshot hits restore selected views, not a tested guarantee of current publisher bytes. HTTP cache and htmx history storage are different mechanisms. Vary names both HX-Request and HX-History-Restore-Request because both influence the fixture's representation. The fixture's historyRestoreAsHxRequest=false is explicit, not a claim about defaults.

The first run completed ten Chromium rows, then failed to locate Firefox in its default installation path. Its incomplete receipt is preserved. The second run used the existing /tmp/fep-playwright-browsers installation and unrestricted temporary profile access. No source/criteria changes between runs; repetition is not new learning. No unexpected assertion failure was discarded. Initial launch error: executable absent at /home/worsch/.cache/ms-playwright/firefox-1543/firefox/firefox.

Learning disposition: corroboration

Before: documentary evidence and synthetic reasoning predicted native contracts, history recovery and representation-aware caching were necessary for coherent routes. After: one pinned library plus our server policy supplies bounded runtime support, including a concrete fragment-as-document counterexample when cache selection is broken. The existing obligations stand; no new pattern or consequential conceptual revision is claimed. Updating evidence text is not automatically a new discovery. This is one question/outcome, not twenty discoveries, and not independent deployed recurrence.

CAND-0009 and CAND-0010 remain draft/hypothesis. VC-03 remains provisional: its tested read-only navigation subset is stronger, while keyboard/focus/announcements, delayed script activation, interrupted transport, invalid/deleted routes, numerical delivery budgets and offline availability remain open. Native recovery was explicitly offered by the fixture; no automatic fallback or full accessibility claim follows. No real CDN/intermediary, private cache policy or publisher-freshness guarantee was tested.

Measurement and next decision

Effort receipt records the monotonic measured window and subtracted execution intervals. Active time includes registration, fixture development, source inspection, analysis and recording. The initial orientation is excluded; subprocess execution includes the failed launch and full rerun. Administrative commit/snapshot/sync time after cutoff is unmeasured overhead. This is single-agent elapsed effort, not human labor or model compute time. Token/cost telemetry is unknown. Brief checkpoint inspection during execution is assigned wholly to execution, never counted twice. Clock start UTC label was read during orientation; monotonic start after orientation is the duration authority.

No diminishing-returns claim follows from one corroborating cycle. WP-0006 T03 keeps comparison-window review and review-date reconciliation open. WP-0003 T01 next owns bounded keyboard/focus/error-feedback checks if we continue this navigation stream; use a new comparison group if the accessibility question/protocol is materially different. WP-0003 T02 still owns the independent walkthrough review and offline branch. M1/M2, independent human review and language usefulness remain unresolved.