ARX4 ideation: Discord packing bets (explore #2 implemented)

docs/arx4-bet2-bench.md

@@ -0,0 +1,108 @@
+# ARX4 bet #2 — content-first / CBOR + real Brotli shared dictionary
+
+_Experimental bench from `scripts/bench-arx4-bet2.mjs`. Not a shipped codec._
+
+## What was implemented
+
+1. **Content-first binary envelope** — `src/lib/payload/arx4-content-first.ts`
+   - Wire: `A4 | version | kind | id | content | meta`
+   - Round-trip tested; stamps rebuilt envelopes as `codec: "plain"` (ARX4 is not shipped).
+2. **CBOR-ish tuple encoder** — same module; encodes the ARX2/3 tuple without JSON quotes.
+3. **Real Brotli shared dictionary** — measured via system `brotli -D` (LZ77 raw dictionary).
+   - Node `zlib.brotliCompressSync({ dictionary })` **silently ignores** the option on Node 22.
+   - Product `brotli-wasm@3` has **no** custom-dictionary API.
+   - Residual `brotli(dict‖data)−brotli(dict)` is kept only as a calibration column.
+
+## Environment
+
+- Node: `v22.14.0`
+- Brotli CLI available: **yes**
+- Shared dictionary size: **1143** bytes (ARX slot text + scaffolding)
+- Discord framing budget (current host): **1962** payload chars
+
+## Pre-compress sizes (envelope bytes before Brotli)
+
+| Fixture | ARX3 substituted | CBOR | CBOR+text-sub | Content-first | CF+text-sub |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| markdown-agents | 6871 | 8056 | 7006 | 8050 | 7000 |
+| code-bench-report | 8203 | 8373 | 8360 | 8366 | 8353 |
+| code-fragment | 7278 | 8065 | 7037 | 8063 | 7070 |
+| json-package | 346 | 308 | 294 | 307 | 294 |
+| small-markdown | 183 | 217 | 186 | 225 | 204 |
+
+## Brotli bytes (q11) vs ARX3
+
+| Fixture | raw chars | ARX3 (baseline) | ARX3 + residual dict est. | ARX3 + real Brotli −D | ARX3 deflate+dict (non-goal) | CBOR tuple + Brotli | CBOR + text-sub + Brotli | CBOR + real Brotli −D | CBOR + text-sub + real −D | Content-first + Brotli | Content-first + text-sub + Brotli | Content-first + real −D | Content-first + text-sub + real −D | Best real bet #2 candidate |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| markdown-agents | 8,000 | 401 (0.0%) | 351 (−12.5%) | 380 (−5.2%) | 471 (+17.5%) | 392 (−2.2%) | 397 (−1.0%) | 385 (−4.0%) | 379 (−5.5%) | 399 (−0.5%) | 403 (+0.5%) | 390 (−2.7%) | 387 (−3.5%) | 379 (−5.5%) |
+| code-bench-report | 8,314 | 2219 (0.0%) | 2188 (−1.4%) | 2219 (0.0%) | 2624 (+18.3%) | 2208 (−0.5%) | 2209 (−0.5%) | 2202 (−0.8%) | 2206 (−0.6%) | 2217 (−0.1%) | 2222 (+0.1%) | 2217 (−0.1%) | 2219 (0.0%) | 2202 (−0.8%) |
+| code-fragment | 8,000 | 308 (0.0%) | 272 (−11.7%) | 307 (−0.3%) | 360 (+16.9%) | 300 (−2.6%) | 296 (−3.9%) | 314 (+1.9%) | 290 (−5.8%) | 311 (+1.0%) | 310 (+0.6%) | 317 (+2.9%) | 299 (−2.9%) | 290 (−5.8%) |
+| json-package | 259 | 182 (0.0%) | 141 (−22.5%) | 188 (+3.3%) | 184 (+1.1%) | 173 (−4.9%) | 173 (−4.9%) | 177 (−2.7%) | 177 (−2.7%) | 190 (+4.4%) | 182 (0.0%) | 181 (−0.5%) | 183 (+0.5%) | 177 (−2.7%) |
+| small-markdown | 189 | 141 (0.0%) | 110 (−22.0%) | 143 (+1.4%) | 160 (+13.5%) | 160 (+13.5%) | 140 (−0.7%) | 144 (+2.1%) | 142 (+0.7%) | 184 (+30.5%) | 155 (+9.9%) | 177 (+25.5%) | 154 (+9.2%) | 142 (+0.7%) |
+
+## Visible chars @ baseBMP vs ARX3
+
+| Fixture | ARX3 (baseline) | ARX3 + residual dict est. | ARX3 + real Brotli −D | ARX3 deflate+dict (non-goal) | CBOR tuple + Brotli | CBOR + text-sub + Brotli | CBOR + real Brotli −D | CBOR + text-sub + real −D | Content-first + Brotli | Content-first + text-sub + Brotli | Content-first + real −D | Content-first + text-sub + real −D | Best real bet #2 candidate | fits Discord? |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | :---: |
+| markdown-agents | 205 (0.0%) | 180 (−12.2%) | 194 (−5.4%) | 240 (+17.1%) | 200 (−2.4%) | 203 (−1.0%) | 197 (−3.9%) | 194 (−5.4%) | 204 (−0.5%) | 206 (+0.5%) | 199 (−2.9%) | 198 (−3.4%) | 194 (−5.4%) | yes |
+| code-bench-report | 1119 (0.0%) | 1103 (−1.4%) | 1119 (0.0%) | 1322 (+18.1%) | 1113 (−0.5%) | 1114 (−0.4%) | 1110 (−0.8%) | 1112 (−0.6%) | 1118 (−0.1%) | 1120 (+0.1%) | 1118 (−0.1%) | 1119 (0.0%) | 1110 (−0.8%) | yes |
+| code-fragment | 158 (0.0%) | 140 (−11.4%) | 158 (0.0%) | 184 (+16.5%) | 154 (−2.5%) | 152 (−3.8%) | 161 (+1.9%) | 149 (−5.7%) | 160 (+1.3%) | 159 (+0.6%) | 163 (+3.2%) | 154 (−2.5%) | 149 (−5.7%) | yes |
+| json-package | 95 (0.0%) | 74 (−22.1%) | 98 (+3.2%) | 96 (+1.1%) | 90 (−5.3%) | 90 (−5.3%) | 92 (−3.2%) | 92 (−3.2%) | 99 (+4.2%) | 95 (0.0%) | 94 (−1.1%) | 95 (0.0%) | 92 (−3.2%) | yes |
+| small-markdown | 74 (0.0%) | 59 (−20.3%) | 75 (+1.4%) | 84 (+13.5%) | 84 (+13.5%) | 74 (0.0%) | 76 (+2.7%) | 75 (+1.4%) | 96 (+29.7%) | 81 (+9.5%) | 92 (+24.3%) | 81 (+9.5%) | 75 (+1.4%) | yes |
+
+## Totals (all fixtures)
+
+| Variant | Σ brotli | vs ARX3 | Σ BMP chars | vs ARX3 BMP |
+| --- | ---: | ---: | ---: | ---: |
+| ARX3 (baseline) | 3251 | 0.0% | 1651 | 0.0% |
+| ARX3 + residual dict est. | 3062 | −5.8% | 1556 | −5.8% |
+| ARX3 + real Brotli −D | 3237 | −0.4% | 1644 | −0.4% |
+| ARX3 deflate+dict (non-goal) | 3799 | +16.9% | 1926 | +16.7% |
+| CBOR tuple + Brotli | 3233 | −0.6% | 1641 | −0.6% |
+| CBOR + text-sub + Brotli | 3215 | −1.1% | 1633 | −1.1% |
+| CBOR + real Brotli −D | 3222 | −0.9% | 1636 | −0.9% |
+| CBOR + text-sub + real −D | 3194 | −1.8% | 1622 | −1.8% |
+| Content-first + Brotli | 3301 | +1.5% | 1677 | +1.6% |
+| Content-first + text-sub + Brotli | 3272 | +0.6% | 1661 | +0.6% |
+| Content-first + real −D | 3282 | +1.0% | 1666 | +0.9% |
+| Content-first + text-sub + real −D | 3242 | −0.3% | 1647 | −0.2% |
+| Best real bet #2 candidate | 3190 | −1.9% | 1620 | −1.9% |
+
+## Findings
+
+### Content-first / CBOR alone
+
+- Dropping JSON (CBOR) or skipping the tuple (content-first) changes pre-Brotli size, but
+  **after Brotli q11 the win vs ARX3 is small** — often within ~1%, and content-first alone
+  can *lose* on small fixtures because it skips ARX text substitution.
+- Re-applying v1 text substitution *inside* content-first / CBOR fields recovers most of
+  that gap; see the `+ text-sub` columns.
+
+### Real Brotli shared dictionary (`brotli -D`)
+
+- Unlike the residual estimate (often −10% to −30%), **real LZ77 shared dictionaries are
+  modest** on this corpus (typically under ~1% total, fixture-dependent).
+- Residual estimates systematically **overstate** the win; do not use them as a ship gate.
+- Deflate+dictionary remains a **non-goal**: larger than plain Brotli on this corpus.
+
+### Product implications
+
+1. **Browser path blocked for real shared dicts today** — `brotli-wasm` has no dictionary
+   API; Node zlib ignores `dictionary`. Shipping ARX4 shared-dict needs a wasm fork or
+   alternate compressor with custom-dict support.
+2. **Binary envelopes are still useful plumbing** (no JSON escaping, cleaner wire) but are
+   not a Discord capacity unlock by themselves on this corpus.
+3. Prefer measuring with **real `brotli -D`** (or a dict-capable wasm) over residual proxies
+   before committing to a shared-dictionary protocol.
+4. Next exploration should focus on **dict contents matched to the post-substitution byte
+   stream** (or a wasm dict path), not more residual optimism.
+
+## How to re-run
+
+```bash
+# requires system `brotli` CLI for real −D columns (apt install brotli)
+npm run bench:arx4-bet2
+# or: node scripts/bench-arx4-bet2.mjs
+```
+
+_Generated in 292.4ms._

docs/arx4-ideation.md

@@ -0,0 +1,222 @@
+# ARX4 ideation — squeezing more into a Discord message
+
+_Experimental notes from `scripts/arx4-ideation-probe.mjs`. Not a shipped codec._
+
+## Goal
+
+ARX3 already optimizes **visible fragment characters** via baseBMP (~15.92 bits/char).
+Discord's hard limit is the full markdown link:
+
+```text
+[label](https://host/path#<tag><payload>)  ≤  2000 characters
+```
+
+So ARX4 should optimize **Discord markdown-link length**, not just fragment length.
+
+## Discord budget math
+
+| Framing | Framing overhead | Payload budget | Max brotli bytes @ baseBMP | Max @ baseAstral (code-point count) |
+| --- | ---: | ---: | ---: | ---: |
+| currentHost (`[Artifact](https://agent-render.com#c…)`) | 38 | 1962 | 3898 | 4897 |
+| shortHost (`[a](https://arx.page#d…)`) | 23 | 1977 | 3928 | 4934 |
+| bareHost (`[x](https://r.page#d…)`) | 21 | 1979 | 3932 | 4939 |
+
+Takeaway: host + label overhead is only ~20–50 chars (and short labels are already the
+agent skill default). The real ceiling is ~3.8–3.9 KB of compressed bytes under baseBMP.
+Astral density (~4.9 KB) is not a Discord win under UTF-16 counting — see research note.
+
+## Wire density ceiling
+
+| Encoding | Bits / JS `length` unit | Notes |
+| --- | ---: | --- |
+| base64url | 6.00 | ASCII-safe; ARX2 default on hostile surfaces |
+| base76 | 6.27 | ASCII fragment-safe |
+| base1k | 10.79 | BMP subset |
+| baseBMP (ARX3) | 15.92 | Current best visible density |
+| baseAstral (code points) | 20.00 | ~26% denser **only if** Discord counts code points |
+| baseAstral (UTF-16 units) | 10.00 | **Worse** than baseBMP (~10 vs ~15.92 bits/unit) |
+
+### Research note — Discord length counting (2026-07 web pass)
+
+Public sources disagree; the **client-facing** signal matters most for paste-to-send:
+
+| Source | Claim | Weight |
+| --- | --- | --- |
+| TypeCount / Discord character-counter guides | Standard emoji count as **2** toward the 2000 limit ("Unicode encoding") | High for UX — matches JS/Electron `.length` |
+| Discord desktop stack | Electron → JS strings → UTF-16 code units | High — composer counter almost certainly uses this |
+| Our product (`markdown-link.ts`) | Already gates on `markdownLink.length` (UTF-16 units) | Aligns with client-side folk wisdom |
+| twilight-interactions #41 | Slash-command option `min/max_length` uses Unicode **code points** (Python `len()`), not UTF-8 bytes | Medium — different API surface than message `content` |
+| Secondary blogs (go-tools, discord-webhook) | "Code points, emoji = 1" | Low — contradicted by emoji=2 guides; some validators still use JS `.length` |
+
+**Working conclusion:** treat Discord message limits as **UTF-16-unit** (JS `.length`) until a live paste test proves otherwise.
+That **kills baseAstral as a Discord win** — astral scalars cost 2 units each, so density drops below baseBMP.
+Keep a one-shot paste test on the backlog (1999 BMP chars vs 1000 astral + framing) only to close the API-vs-client gap; do not prototype baseAstral for Discord first.
+
+## Ideas probed
+
+1. **baseAstral wire** — pack into supplementary-plane scalars (~20 bits/code point). **Deprioritized for Discord** after UTF-16 research.
+2. **CBOR-ish binary tuple** — drop JSON quotes/escapes around the ARX2/3 tuple. **Worth exploring.**
+3. **Brotli shared static dictionary** — seed Brotli with domain patterns already in `/arx-dictionary.json`.
+   Node zlib cannot set a Brotli custom dictionary today, so the probe reports (a) a residual
+   `brotli(dict‖data)−brotli(dict)` estimate and (b) a real `deflateRaw+dictionary` proxy. **Worth exploring** via wasm.
+4. **Content-first binary envelope** — compress raw artifact bytes + tiny binary header, skip JSON entirely. **Worth exploring.**
+5. **Mined overlay growth** — corpus-mined n-grams as an extra substitution layer.
+6. **Combined ARX4 stack** — content-first + mined + shared-dict estimate.
+7. **Discord framing** — short host + 1-char label + compact tag. **Already practiced** (skill uses short labels; host shortening is DNS, not codec).
+
+## Corpus results
+
+Visible char counts assume the ARX3-style compact tag + dense Unicode wire (marker + 2-char length + digits).
+Percentages are vs ARX3 baseBMP visible chars (negative = larger / worse).
+
+### Brotli bytes
+
+| Fixture | raw chars | ARX3 (baseline) | ARX3 + Brotli dict est. | ARX3 deflate+dict proxy | CBOR tuple + Brotli | CBOR + Brotli dict est. | Content-first binary | Content-first + dict est. | ARX3 + mined overlay | Content-first + mined | ARX4 stack (cf+mined+dict) |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| markdown-agents | 8,000 | 401 (0.0%) | 351 (−12.5%) | 471 (+17.5%) | 392 (−2.2%) | 337 (−16.0%) | 395 (−1.5%) | 342 (−14.7%) | 405 (+1.0%) | 398 (−0.7%) | 363 (−9.5%) |
+| code-bench-report | 8,314 | 2219 (0.0%) | 2188 (−1.4%) | 2624 (+18.3%) | 2208 (−0.5%) | 2169 (−2.3%) | 2219 (0.0%) | 2171 (−2.2%) | 2296 (+3.5%) | 2285 (+3.0%) | 2241 (+1.0%) |
+| code-fragment | 8,000 | 308 (0.0%) | 272 (−11.7%) | 360 (+16.9%) | 300 (−2.6%) | 255 (−17.2%) | 308 (0.0%) | 261 (−15.3%) | 322 (+4.5%) | 320 (+3.9%) | 286 (−7.1%) |
+| json-package | 259 | 182 (0.0%) | 141 (−22.5%) | 184 (+1.1%) | 173 (−4.9%) | 129 (−29.1%) | 184 (+1.1%) | 131 (−28.0%) | 190 (+4.4%) | 192 (+5.5%) | 150 (−17.6%) |
+| small-markdown | 189 | 141 (0.0%) | 110 (−22.0%) | 160 (+13.5%) | 160 (+13.5%) | 103 (−27.0%) | 163 (+15.6%) | 108 (−23.4%) | 154 (+9.2%) | 178 (+26.2%) | 118 (−16.3%) |
+
+### Visible chars @ baseBMP
+
+| Fixture | ARX3 BMP | best idea BMP | win | fits Discord (current host) | fits (short host) |
+| --- | ---: | ---: | ---: | :---: | :---: |
+| markdown-agents | 205 | 173 (CBOR + Brotli dict est.) | −15.6% | yes → yes | yes |
+| code-bench-report | 1119 | 1093 (CBOR + Brotli dict est.) | −2.3% | yes → yes | yes |
+| code-fragment | 158 | 132 (CBOR + Brotli dict est.) | −16.5% | yes → yes | yes |
+| json-package | 95 | 68 (CBOR + Brotli dict est.) | −28.4% | yes → yes | yes |
+| small-markdown | 74 | 55 (CBOR + Brotli dict est.) | −25.7% | yes → yes | yes |
+
+### Visible chars @ baseAstral (optimistic code-point counting)
+
+| Fixture | ARX3 BMP | ARX3 Astral | best idea Astral | astral win vs ARX3 BMP |
+| --- | ---: | ---: | ---: | ---: |
+| markdown-agents | 205 | 164 | 138 (CBOR + Brotli dict est.) | −32.7% |
+| code-bench-report | 1119 | 891 | 871 (CBOR + Brotli dict est.) | −22.2% |
+| code-fragment | 158 | 127 | 106 (CBOR + Brotli dict est.) | −32.9% |
+| json-package | 95 | 76 | 55 (CBOR + Brotli dict est.) | −42.1% |
+| small-markdown | 74 | 60 | 45 (CBOR + Brotli dict est.) | −39.2% |
+
+### Totals (fixtures with a value for that variant)
+
+| Variant | Σ brotli | Σ BMP chars | vs ARX3 BMP | Σ Astral chars | vs ARX3 BMP |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| ARX3 (baseline) | 3251 | 1651 | 0.0% | 1318 | −20.2% |
+| ARX3 + Brotli dict est. | 3062 | 1556 | −5.8% | 1243 | −24.7% |
+| ARX3 deflate+dict proxy | 3799 | 1926 | +16.7% | 1538 | −6.8% |
+| CBOR tuple + Brotli | 3233 | 1641 | −0.6% | 1312 | −20.5% |
+| CBOR + Brotli dict est. | 2993 | 1521 | −7.9% | 1215 | −26.4% |
+| Content-first binary | 3269 | 1660 | +0.5% | 1326 | −19.7% |
+| Content-first + dict est. | 3013 | 1531 | −7.3% | 1223 | −25.9% |
+| ARX3 + mined overlay | 3367 | 1709 | +3.5% | 1365 | −17.3% |
+| Content-first + mined | 3373 | 1713 | +3.8% | 1368 | −17.1% |
+| ARX4 stack (cf+mined+dict) | 3158 | 1605 | −2.8% | 1282 | −22.4% |
+
+## Discord capacity (how much raw text fits)
+
+Binary-searching the largest source string whose encoded markdown link stays ≤ 2000 chars
+(current-host framing, payload budget 1962) shows the practical ceiling:
+
+| Content shape | ARX3 @ baseBMP | ARX3 @ baseAstral | Content-first @ BMP | Notes |
+| --- | ---: | ---: | ---: | --- |
+| Tiled real report + unique headers | ≥120k (search cap) | ≥150k (~+25%) | ≥120k | Highly compressible; Discord is not the bottleneck |
+| Generated TS helpers | ≥120k (search cap) | ≥150k (~+25%) | ≥120k | Same |
+| Quote/newline-heavy prose | ~88k | ~112k (~+27%) | ~87k | JSON escaping barely matters once Brotli runs |
+| Current `code-bench-report.md` (8.3k) | ~1117 BMP chars | ~891 astral | ~1117 | Uses ~57% of Discord budget today |
+
+**Reading:** for Discord, ARX3 already leaves a lot of headroom on typical artifacts. With
+baseAstral deprioritized (UTF-16), the ways to *raise the ceiling* are fewer compressed
+bytes via shared dictionaries / better envelopes — not denser Unicode wire or more text substitutions.
+
+## Interpretation
+
+### What already works in ARX3
+
+- For typical single artifacts under ~8–12 KB of source, ARX3 baseBMP already fits Discord
+  with room to spare (see `small-markdown`, `json-package`, `code-fragment`, and the
+  8.3k code-bench report at ~57% of budget).
+- The painful case is large unique prose/reports where dictionary substitution helps less
+  and Brotli carries most of the work — chase bytes there, not astral wire.
+
+### Ranked bets for ARX4
+
+1. **Agent-side semantic splits (skill) — preferred over mosaic**
+   - When Discord budget blows, agents should cut by meaning (sections/files) into
+     independent links — see `skills/agent-render-linking/SKILL.md`.
+   - **Mosaic/`1/N` reassembly is out** — not worth the protocol/UX cost.
+
+2. **Kind-specific dictionaries — measured, not yet (see `docs/arx4-kind-dicts.md`)**
+   - Free kind tags (`m`/`k`/`j`/…) cost **0** extra chars vs today’s `c`; a +1 selector is optional.
+   - On this LOO corpus, kind overlays **do not beat** shared ARX3 (Σ BMP **+0.4%** free-tag,
+     **+0.8%** with +1 selector). Wins were tiny/rare (2/7 fixtures free-tag).
+   - **Do not spend a char hoping to “save thousands”** — Discord wins look like tens of BMP
+     chars when they exist; agents should split oversized source instead.
+   - Revisit only with a larger per-kind held-out corpus.
+
+3. **Content-first / CBOR + real Brotli shared dict — explored (see `docs/arx4-bet2-bench.md`)**
+   - Residual dict estimates overstated wins (~6–8% corpus); **real `brotli -D` is ~0–1%**.
+   - Binary envelopes alone are not a Discord unlock; keep as plumbing.
+
+4. **Versioned shared prior (chunk / word) — only with LOO gate**
+   - Warm priors are fake (−70%+); LOO chunk priors **regressed** (~+20% BMP). See silly-cuts.
+
+5. **Implied envelope + kind IR + lossy — small honest cuts**
+   - Implied / kind IR ~3–4% corpus; lossy ~8% when stripping emphasis/comments.
+
+6. **Hybrid fence / label bitstream — wash**
+
+7. **Curated overlay growth (cautious)** — mined n-grams alone regressed (+3–4%).
+
+8. **baseAstral — deprioritized for Discord** (UTF-16 counting).
+
+9. **Discord framing — already practiced, not an ARX4 lever**
+
+### Suggested ARX4 shape (if pursued)
+
+```text
+# Single-link path (marginal):
+artifact bytes
+  → implied / content-first envelope + optional kind IR
+  → optional kind-tagged dict ONLY if held-out gate clears (prefer free tags m/k/j/…)
+  → Brotli q11 (+ real shared dict only if browser wasm gains dict support)
+  → baseBMP → compact tag
+
+# Over-budget path (product, not codec):
+agent semantically splits → N independent ARX3 links (skill guidance)
+```
+
+Selection policy: optimize `markdownLink.length` (JS/UTF-16 units) for a declared surface
+(`discord` | `visible` | `transport`).
+
+Takeaway: **stop mosaic; stop paying selector chars without a LOO win; prefer skill splits
+and only revisit kind dicts with a bigger corpus.**
+
+## Non-goals / traps
+
+- Do not weaken the 8192 fragment budget or 200k decoded budget for Discord wins.
+- Do not put artifact bodies in query params.
+- Do not chase baseAstral for Discord until a live paste test overturns UTF-16 counting.
+- Do not treat short-host framing as an ARX4 deliverable.
+- Do not replace UUID mode: hostile link scanners still want short opaque URLs.
+- Do not grow substitution dictionaries without a corpus gate — mining can regress.
+
+## How to re-run
+
+```bash
+npm run bench:arx4-ideation
+# or: node scripts/arx4-ideation-probe.mjs
+
+# Bet #2 follow-up (content-first/CBOR + real brotli -D):
+npm run bench:arx4-bet2
+
+# Silly / deep cuts (priors, IR, mosaic math — mosaic deprioritized):
+npm run bench:arx4-silly
+
+# Kind-specific dictionaries (+ free tag vs +1 selector):
+npm run bench:arx4-kind-dicts
+```
+
+_Generated in 9.6ms._
+See also `docs/arx4-bet2-bench.md`, `docs/arx4-silly-cuts.md`, and `docs/arx4-kind-dicts.md`.