Remove reference-game replication experiments (ported to their own repo)

examples/README.md

@@ -646,52 +646,3 @@ and the matcher reproduces that order by assigning each shape to a slot, then su
 out of N. Shows that "sequential" and "unconstrained" are one plugin with `targets` turned from
 length-1 to length-N. Run it the same two-tab way.
 
-## `reference-game-hawkins.html` and `reference-game-cwg.html`
-
-Two **faithful replication** builds of the reference game, distinct from the two demos above: where
-`reference-game.html` / `reference-game-match.html` are lightweight teaching demos (synthetic shapes,
-short schedules), these two are **study artifacts** parameterised to reproduce a specific published
-paper, sharing the canonical 12-tangram stimulus set via [`tangrams.js`](tangrams.js). The tangrams in
-[`assets/tangrams/`](assets/tangrams/) are the exact set from
-[hawkrobe/tangrams](https://github.com/hawkrobe/tangrams) (Hawkins, Frank & Goodman 2020), themselves
-reproduced from Clark & Wilkes-Gibbs (1986) — so both builds are directly comparable to both papers.
-
-| | `reference-game-hawkins.html` | `reference-game-cwg.html` |
-| --- | --- | --- |
-| Replicates | Hawkins, Frank & Goodman (2020), *Cognitive Science* 44 | Clark & Wilkes-Gibbs (1986), *Cognition* 22:1–39 |
-| Condition | Sequential — **single** target, one click | Full-board — all **12** tangrams as an ordered target |
-| Schedule | 6 blocks × 12 = **72 trials** (`sequentialSchedule`) | **6 trials**, fresh full-board order each (`fullBoardSchedule`) |
-| Fidelity note | Near-exact (the paper was already web + text chat), including Exp. 2's matcher-click gate | Text chat **substitutes** for the original **spoken** dialogue — the one deliberate deviation, noted in-file |
-
-Both fix director/matcher for the whole game, reveal the target to the director only ("cued"), and use
-**unrestricted two-way chat** (`chat_role: "both"`, no count/length limits) — faithful to both papers,
-and load-bearing for the reference-shortening effect. `tangrams.js` builds each schedule
-**deterministically** from a seed, and both files seed from the shared session id (`?mp_session=`):
-each dyad therefore gets its own trial order while its two partners still agree on it without
-exchanging anything — a single fixed order across dyads would confound item order with repetition
-block, which is the axis both papers measure. Both files also **preload** the 12 tangram PNGs before
-pairing, since `rt` is a reported DV and `stimuli` is an object parameter that jsPsych's automatic
-media preloading can't see into. The faithful schedules are long for manual piloting — lower `BLOCKS`
-(Hawkins) / `TRIALS` (C&WG) at the top of each file for a quick smoke test, then restore.
-
-Both configs were verified parameter-by-parameter against the papers **and** the original
-[hawkrobe/tangrams](https://github.com/hawkrobe/tangrams) experiment code (which pins the details the
-prose leaves ambiguous — e.g. the code confirms director and matcher grids are *independently*
-scrambled, `notMatchingLocs`, and that the schedule is 72 rounds / 6 per tangram for the sequential
-build and `numRounds = 6` full boards for the unconstrained build). Both have been play-tested two-tab
-at full faithful schedule length, collecting both participants' data.
-
-The Hawkins build sets `require_message_before_response: true`, reproducing Exp. 2's gate on the
-matcher's click (`game.client.js`: `if (globalGame.messageSent)`) so a referring expression exists on
-every trial. The C&WG build leaves it off, matching its free-form protocol.
-
-> **`round_timeout` is not from either paper.** Both builds set one (60s Hawkins, 180s C&WG) purely so
-> a partner who disconnects or walks away cannot hang the trial forever — neither original was timed.
-> Rounds it ends are logged as `ended_by: "timeout"` with a null assignment, so they are easy to
-> exclude; raise the values if piloting shows genuine trials running long, especially C&WG's first
-> full-board trials, which are the longest in the study.
-
-Like the demos they run on `adapter-multiplayer-local` for two-tab piloting; the header comment marks
-the **one-line swap to `adapter-multiplayer-firebase`** (plus a real waiting room) for the paid
-Prolific run. Run them the same two-tab way (build packages, serve the repo, copy the `?mp_session=…`
-URL into a second tab).