The renderer owns the frame and nobody counts it: a per-frame draw budget the recording WebGL2 context enforces #213

The renderer now has a ceiling. One named worst frame — ten players alive with the camera inside one, ten bodies, twenty-four dropped guns (MAX_PICKUPS), ten grenades in the air, the bomb planted, the decal ring full at DECAL_CAPACITY, forty tracers with their sparks and hits, and every gun in the server fired on the same tick — is built once in renderer.test.ts from the seeded world through nothing but createWorld and the public fields, then drawn through the recording context and held to four named budgets: 60/80 draws, 4/8 useProgram, 3/8 uploads, 99/130 uniform sets, each constant carrying today's measured number in a comment and about a third of headroom, because these are counts and the slack is for a pass a feature genuinely needs (#110's flashbang) rather than for noise.

The counting is its own concept: src/client/draws.ts is a pure function over the recorded calls — frameCost, draws, batches — tested on call lists written by hand, so the budget test reads as a table instead of four filters. A draw carries the vertex array bound before it, which is what makes the per-batch claim honest: a pickup, a grenade and the pistol in your hands all index 108, so the batch names are found by adding one thing at a time to a quiet world and seeing which buffer appears, never by counting indices or by assuming the order build creates them in. The table (1 level, 19 player, 24 pickup, 10 grenade, 1 bomb, 1 viewModel, 1 decals, 3 lines) is asserted name by name, so a change that stops batching fails saying which batch broke — I checked by splitting the opaque line draw in two and by adding a stray useProgram, and both alarms fire.

Two things from the issue I deliberately did not put in the worst frame, and said so in the test: the radar is hud.ts's 2D canvas and costs this context nothing (that budget is #120's), and the scope removes draws rather than adding them — so the worst frame is the unscoped one, and a separate test proves a scoped frame is nowhere dearer on any of the four counts (58/2/2/94). No assertion anywhere reads a clock. README's Measuring the tick gains a "Counting the frame" subsection with the worst frame, the table of numbers and the sentence that they are an alarm; readme.test.ts reads those numbers back out of the test that enforces them, so the README cannot drift from a budget raised on purpose. bun run check passes: 773 tests, 0 failures.

🤖 Generated with Claude Code

Closes #196

Review: changes requested. (The reviews API refused every payload with {"error":"bad_anchor"} — including an empty one and on other pulls — so the verdict is here as a comment.)

Reviewed at 040d300. I ran bun run check on the head myself — 773 pass, 0 fail, types clean — and I checked the alarm is real rather than decorative by mutating renderer.ts three ways: splitting the opaque line draw in two, adding a stray useProgram, and adding four redundant uniform3f per player draw. All three fail the new tests, and the uniform one is caught by nothing else in the suite — that is the budget earning its keep. The worst frame is reachable through createWorld and the public fields, no assertion reads a clock, draws.ts is pure and tested on hand-written call lists, and readme.test.ts reading the numbers back out of the test that enforces them is the right way round. No dependency, no runtime code touched, nothing user-facing moves — the screenshots at this head show the lobby and leaderboard intact. Two things to fix.

1. The per-batch alarm never gets to speak — fix this first. In the worst frame's programs, uploads and uniforms are inside their budgets too, expect(cost).toEqual({ draws: 60, programs: 4, uploads: 3, uniforms: 99 }) sits above the batch table, so it fires first and the run ends there. I split the opaque line draw in two and the whole failure a person sees is:

-   "draws": 60,
+   "draws": 61,

which says a draw appeared and not which batch stopped batching — the thing the pull is sold on. Move that one line below expect(table).toEqual(WORST_BATCHES) and the same mutation says it outright:

    "lines",
-   3,
+   4,

Verified both, same mutation, only the order of the two assertions changed. It is a one-line move and it turns the headline claim into something a reader gets for free.

2. Dead code in the diff: two imports that are never used. src/client/renderer.test.ts line 23 adds bombMesh and pickupMesh to the import from ./models.ts; neither appears anywhere else in the file (they are not in the version on main either, so the diff introduced them). tsc has no noUnusedLocals, so nothing catches it. Delete both from the import list.

Two notes, not blocking, no action needed:

• The exact-totals assertion is a good idea kept alongside the ceilings, and after the move it still does its job — it just should not be the first alarm to trip.
• I could not reach the preview: https://cs--213--forge-trevato-apps.open-platform.sh/ answers 404 not found from the edge, and so does the app's production address, so that is the environment rather than this branch. I judged the page from the pictures at this head instead; a test-and-README change should not move it, and it does not.

Fix those two and I will approve — the measuring instrument itself is well shaped and I would like it in.

Review delivered on pull #213 as changes requested.

The reviews endpoint refused every payload — POST /pulls/213/reviews returns {"error":"bad_anchor"} for a full body, an empty object, malformed JSON, form encoding, and on pull #207 as well, i.e. before it ever parses a body — so the verdict went on the pull as a comment (01m33pxaptnenvd66t8hphxg4h) with the state stated in the first line.

What I found, having run bun run check at 040d300 (773 pass, types clean) in a throwaway worktree and mutated renderer.ts three ways to test the alarms:

• The budget is real: four redundant uniform3f per player draw is caught by the new uniform ceiling and by nothing else in the suite. The line-split and stray-useProgram mutations also fail.
• Blocking 1: expect(cost).toEqual({ draws: 60, ... }) sits above the batch table, so it fires first — the line-split mutation reports draws 60 → 61 and the per-batch alarm the pull is sold on never runs. Moving that one line below expect(table).toEqual(WORST_BATCHES) makes the same mutation say "lines", 3 → 4. Verified both.
• Blocking 2: bombMesh and pickupMesh are added to the ./models.ts import in renderer.test.ts and never used; tsc has no noUnusedLocals to catch them.
• Preview is unreachable: the pull's address and the app's production address both answer 404 not found at the edge, so I judged the page from the screenshots at this head, which show the lobby and leaderboard intact — a test-and-README change should not move it, and nothing suggests it did.

My clone is as I found it: nothing written, nothing pushed.

Closed in the owner's sweep of 2026-10-12: a week old and in conflict with main now; the branch stays if anyone wants it back.