Preregister first measured navigation research cycle

Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a0cffe-206f-7323-89d3-b20ca4cf4bf9

research/evolution/cycles/2026-09-27-navigation.md

@@ -0,0 +1,77 @@
+# LEARN-0011 — Navigation representations across entry and history
+
+Status: preregistered before implementation/execution; owner/reviewer Codex (author-model).
+Local date 2026-09-27 (Europe/Berlin); UTC dates may still be 2026-09-26.
+Stream B; CAND-0009 URL-addressable State and CAND-0010 Progressive Enhancement.
+
+## Question and prior belief
+
+Can one public read-only route preserve selection and complete-document availability
+across direct entry, absent/failed enhancement, history restoration and HTTP caching?
+Prior: a native href/GET contract plus representation-aware routing/cache selection
+should satisfy this narrow obligation. Merely pushing a URL, emitting HTML or adding
+Vary without checking actual cache behavior is insufficient. History snapshots may
+restore older bytes; no freshness or offline availability promise follows.
+
+Expected decision: corroborate these existing obligations with a pinned runtime, or
+revise their scope if the failure cases expose an unrecorded obligation. A pass count
+alone is not a revision. One question yields one ledger outcome, not one per test.
+Disconfirmation: a correctly keyed full/fragment response is mixed; a valid native
+contract needs client script; or the tested history recovery cannot resolve selection
+under the declared setup. Unexpected results remain in the receipts.
+
+## Registered comparison contract: B-navigation-mechanism-v1
+
+One bounded route/representation question per cycle; real pinned library in authored
+local HTTP fixture; Chromium and Firefox; positive paths plus negative controls;
+author-model review, no production recurrence or human validation inferred. Use this
+group only for comparable question size/evidence tier and the same accounting below.
+No model inference is needed. Any later inference must use llm-connect.
+
+Pin htmx 2.0.11 and Playwright 1.63.0 with lockfile. Use native anchors and a GET filter,
+optional boosted links with explicit history configuration. Keep application server
+policy separate from vendor behavior. Document exact headers, browser versions,
+file hashes and request traces. Real browser HTTP cache, no request interception.
+
+## Checks fixed before execution
+
+Run each browser condition once, isolated contexts; retain failures, do not retry to
+hide them. Wait on events/DOM/network state rather than arbitrary settling sleeps.
+
+| ID | Condition / expected observation |
+| --- | --- |
+| N01 | Direct filtered entry with JavaScript disabled, native link and GET form: full shell, matching selection/URL |
+| N02 | Enhanced link pushes selection; reload and fresh direct entry produce full shell |
+| N03 | Back/forward with history snapshot available: selection restored; distinguish network from snapshot |
+| N04 | Evict htmx history storage before back: history-miss request restores selection from full document |
+| N05 | Enhanced request returns 503: committed old URL/content remain; explicit native recovery can navigate |
+| N06 | htmx script returns 503: native link and GET form still work |
+| N07 | Cache fragment first, then full document for same URL: correct Vary preserves both representations |
+| N08 | Cache full document first, then fragment for same URL: reverse ordering preserves both |
+| N09 | Omit Vary negative control: browser may reuse a fragment as a document; record actual behavior, do not presume immunity if not reproduced |
+| N10 | Validator checks: matching ETag yields 304; other representation ETag yields 200 and correct body; history-restore selector returns full document |
+
+N09 is diagnostic, not a success guarantee: a non-reproduction narrows this browser
+observation and leaves the contract concern open. N10 may use explicit HTTP requests
+for conditional semantics; distinguish these from automatic browser revalidation.
+No real intermediary/CDN, production application, offline guarantee, writes, general
+accessibility or latency claim. Missing parameter/deleted-resource checks, keyboard,
+focus and announcements remain separate obligations unless explicitly added later.
+
+## Effort accounting and budget
+
+Budget: 45 active research minutes, plus at most 10 execution minutes before a scope
+review. Single-agent continuous session. A monotonic clock starts after orientation,
+before this registration. Active research is elapsed session time minus measured
+blocking experiment/verification subprocess intervals; includes implementation,
+reading, analysis and recording. No concurrent work; no overlapping double count.
+Package setup is research overhead. Pause logging if interrupted. Timer cutoff is the
+completed analysis/verification boundary; final commit/snapshot/sync administration
+is excluded and reported as unmeasured overhead. Receipt-backed tokens/cost unknown;
+no estimate from this chat. Save the timing receipt and state clock precision/limits.
+
+Commit this preregistration before implementation or browser execution. If a harness
+error needs correction, retain the first result and explain the change. Record the
+outcome in a separate review and append-only ledger; this plan stays as registered.
+Remaining work belongs to WP-0003 T01/T02 and WP-0006 T03. One cycle cannot establish
+any learning-rate trend or diminishing returns.