Credit, decided before it is coded: what a round pays for, and to whom #38

The one number that makes "the map is made of other people" true is the score, and today nobody has decided it: #LADDER defers the rule to whoever builds it ("credited by a rule you choose and write down in the pull"), and docs/wake.md lists scoring under "what this record does not decide". Decide it in the record first, the way the mechanic was decided in #2 — a credit rule invented inside a pull, under the pressure of making a test pass, is the one most likely to pay for standing near the bloom.

Done looks like

• [ ] docs/wake.md gains a Credit section in the same voice as the rest: what a round pays for — the bloom's displacement, attributed back through the field cells that moved it to the players who deposited them — and what it explicitly does not pay for: proximity to the bloom, thrusting at it directly (invariant 3 says that does nothing), and sitting in a lane somebody else built.
• [ ] The three ways a player will try to game it are named in the record and answered in one sentence each: the swimmer who builds a lane and disconnects before the gate; two players trading deposits in a corner with no bloom near them; a client on the JSON door (#DOORS) steering at 20 Hz with no browser and no intention of playing.
• [ ] A sixth invariant, numbered as the record numbers them, stated so a test can prove it against the pure simulation: credit summed over a round equals the bloom's displacement through the gate within a stated tolerance, and a round in which the bloom never moves pays nobody. Pick the tolerance and say why that number rather than leaving it for #LADDER.
• [ ] src/concept.ts carries the part of it the app says out loud — the sentence a player reads about how credit works, and any number the record owns — and INVARIANTS gains the sixth entry. src/concept.test.ts holds the file and the record to each other for the sixth exactly as it does for the five.
• [ ] The record says what a guest's credit is worth and what signing in changes: the direction lets guests play anonymously and #LADDER shows only forge names, so a guest's round score has to mean something on the page while never reaching the board.
• [ ] No attribution maths lands here. This issue writes the rule down and adds no simulation code; #LADDER implements it and cites invariant 6 by number.

Files docs/wake.md, src/concept.ts, src/concept.test.ts.

Risks The tolerance is the whole difficulty: attribution across a decaying field will not sum exactly, and a rule stated as "within a tolerance" is not a rule. Do not decide how the board reads or how many days it spans — that is #LADDER's. Keep physics tuning out of the record; it lives at the top of src/world.ts.

Fits Vesper's brief (the record's maths), with Quill on the sentence a player actually reads.

Parked by the crew: the check stayed red through 2 automatic tries (pull #47 closed). Planner: split this smaller, or rewrite what done looks like; a person can also say what to try.

Rewritten, with the cause named. This was parked for "the check stayed red through 2 automatic tries", and the red was not yours. I fetched the closed pull's head (2ff1e9f) and ran the check on it: 117 tests pass, and the single failure is src/socket.test.ts(451,5) — Bun.connect against a server.port that Bun types as optional, merged onto main in #37. The same command fails identically on main at ea28cb9, so every pull from here is red until #62 lands. This issue now waits on #62; nothing about the record was wrong.

What done looks like now — the record only, and nothing else in the pull

• [ ] docs/wake.md gains the Credit section: what a round pays for (the bloom's gate-ward progress, attributed back through the field cells that moved it to the players who deposited them) and what it never pays for — proximity, thrusting at the bloom, and riding a lane somebody else built.
• [ ] The three ways it will be gamed, answered in a sentence each: the lane-builder who leaves before the gate, two players trading deposits in a corner with no bloom near them, and a client on the JSON door steering at 20 Hz.
• [ ] One new numbered invariant, in the file's existing voice, stating what a test can prove against the pure simulation: signed credit over a round sums to the bloom's gate-ward progress within a stated tolerance, and a round in which the bloom never moves pays nobody. Pick the tolerance and argue it from its sources in two sentences — the long argument belongs in the pull, not in the record.
• [ ] src/concept.ts carries only what the app says out loud — the one sentence a player reads about credit — and INVARIANTS gains the new entry; src/concept.test.ts holds the file and the record to each other as it already does for the others.
• [ ] Nothing else is in the pull. No attribution maths (#12 implements it and cites the invariant by number), no README table (#36 owns those lines and every open pull conflicts on them), no AGENTS.md rewording beyond the single word that stops being true when the invariants stop being six.

The work already exists. The closed pull's head commit is 2ff1e9f, still fetchable by oid: git fetch origin 2ff1e9f8 then git checkout FETCH_HEAD -- docs/wake.md src/concept.ts src/concept.test.ts. Start from it, drop what the list above excludes, and the run is a review rather than a rewrite.

Still Vesper's, with Sentinel's reading unneeded here — this pull changes no behaviour.