fix(elision): elide ungated dev-only prose + :doc non-literal + directory sentinel

[mike-thompson-day8]

Summary

Three production-bundle-elision fixes on the shared elision-probe surface (implementation/scripts/check-elision.cjs sentinels + implementation/core/test/re_frame/elision_probe.cljs probe roots). One worker on one branch because all three share the two coupling files. Refs rf2-7czrms, rf2-7xz70g, rf2-tfiutq.

1. Ungated dev-only diagnostic emits — STEP-0 empirical re-evaluation

The ~18 app-facing warning/diagnostic emits (resources / routing / machines / ssr-client) call trace/emit! / trace/emit-error! directly on a runtime data branch (when-some / when (seq …) / doseq / a catch), building (str …) or literal-string prose, with no call-site interop/debug-enabled? gate.

STEP-0 finding (the load-bearing surprise): I rooted a representative site (routing advise-query-promotion! — emit! + (str …) :advice; machines emit-destroy-exit-failure! — emit-error! + literal string) in the :advanced probe and grepped both builds. The prose is already elided in production:

sentinel	control (DEBUG=true)	production (DEBUG=false)
stays a STRING in the route slice	present (1)	absent (0)
An :exit action threw during destroy-time cascade	present (1)	absent (0)

A directly-called emit! / emit-error! folds its body to nil under goog.DEBUG=false; Closure inlines the now-no-op call and drops the pure prose args as dead. The bead's stated root cause — emit!'s identity escaping into a late-bind registry — applies to the framework's own emit machinery (storage.cljc, registrar.cljc, which capture emit! via (late-bind/get-fn :trace/emit!) and are already gated), not to these direct-call sites.

Per the bead's own STEP-0 escape hatch ("if it does NOT survive — already DCE'd — re-evaluate, do not blindly gate"), no mass-gating was applied — gating 18 already-elided sites would add redundant noise implying a leak that does not exist (ELEGANCE / CLARITY). Instead I closed the systemic probe gap the bead correctly identifies: the probe previously pinned dev-only prose absence only inside already-gated branches and via the require boundary — it never rooted an ungated direct-emit site, so the fold-path's production absence was unverified. Added two DEV_ONLY_SENTINELS + a touch-direct-emit-diagnostics! root so the control build exercises the path and a future regression (e.g. capturing emit! into a var, escaping its identity like the framework machinery) is caught.

2. :doc macro-DCE gap on reg-machine opts map

expand-reg-machine forwarded its opts-form verbatim, bypassing gate-doc-args — so a literal (reg-machine :id {:doc "…"} spec) left its :doc string bytes in the :advanced bundle (the splice-through defreg-macro / defreg-event-macro surfaces gate every literal doc-bearing arg; reg-machine did not). Fix: route opts-form through gate-doc-arg when it is a literal map (the idiomatic shape), emitting the same (if interop/debug-enabled? <full> <stripped>) gate Closure folds. A non-literal opts expression (symbol / computed map / merge) passes through unchanged — its :doc still strips from stored handler-meta via the runtime register! strip, but its bytes are outside the macro's reach.

Pinned by: a new DEV_ONLY sentinel + probe touch (rf2-tfiutq-machine-opts-doc-sentinel — prod absent, control present, confirmed); a JVM handler-meta-strip parity test in the machines artefact (re-frame.machine-doc-opts-prod-elision-test); a cross-reference note in the core doc-elision suite (the core :test classpath lacks day8/re-frame2-machines, so the machine surface's semantic strip test lives in the machines artefact).

3. Reachability sentinel for the late-bind directory :description corpus

re-frame.late-bind.directory/hooks carries ~150 paragraph-length :description strings — plain ungated CLJC data, kept out of production solely by reachability DCE (verified: zero production src namespace :requires the directory; only test files and one docstring mention it). No leak today, but no sentinel guarded that absence. Added a PROD_ABSENT_WHEN_UNUSED_SENTINELS entry (Enqueue an event for processing by the drain loop, the :router/dispatch! row's :description, unique across tracked source) mirroring the EP-0023 image-assembly reachability guard, so an accidental future production :require is caught by CI. Pure test-guard addition — no source change.

Gate results

• npm run test:elision — PASS: production 45/45 absent (was 42), control 45/45 present, EP-0023 1/1 provenance-survives + 2/2 assembly-DCE.
• npm run test:cljs — 7999 tests / 36798 assertions, 0 failures, 0 errors.
• machines JVM suite (clojure -M:test) — 845 tests / 2743 assertions, 0 failures, 0 errors.
• core doc-elision JVM test — 5 tests / 27 assertions, 0 failures.
• new machines doc-opts test — 2 tests / 4 assertions, 0 failures.

Files changed

• implementation/core/src/re_frame/core_reg_macros.cljc — route reg-machine literal opts through gate-doc-arg (only production-source change).
• implementation/scripts/check-elision.cjs — +2 direct-emit DEV_ONLY sentinels, +1 reg-machine opts-:doc DEV_ONLY sentinel, +1 directory-corpus PROD_ABSENT_WHEN_UNUSED sentinel.
• implementation/core/test/re_frame/elision_probe.cljs — touch-direct-emit-diagnostics! + reg-machine literal-opts :doc touch.
• implementation/machines/test/re_frame/machine_doc_opts_prod_elision_test.clj — new JVM parity test for the reg-machine opts :doc strip.
• implementation/core/test/re_frame/doc_metadata_prod_elision_test.clj — cross-reference note pointing at the machines-artefact parity test.