Thirty per cent of rounds end on the clock: bots that feel the round timer #137

Approved. Both things I asked for are here, and I checked them the way I checked the first round — by deleting the line and watching the test fall over.

The two gaps are closed.

• Drop if (world.bomb.state === "planted") return false; from pushing and once the bomb is down the clock stops moving the terrorists' feet fails. It is a good test: it puts the carrier in the duel it was already in, plants the bomb under it, and asserts brain.goal stays null, so the thing being pinned is that the policy keeps the feet — not merely that the bot ends up somewhere near the bomb, which chooseGoal would have given you either way.
• Drop :hunt from repath's objective and the defence leaves its angle the tick the clock says so fails. The setup is the right one: it asserts up front that nothing about the bot is stale — part-way through a path, world.time < repathAt — so the only thing that can turn it round is the objective string, and then it moves the clock and nothing else.
• You also closed :called, which was unpinned before this pull ever touched the file: drop it and a call turns a rotating defender round inside its repath window fails. That was not asked for and it is the harder of the three to set up, since the call has to name the site the bot already had for the test to mean anything. Nice.

muster.ct now says in the type why both halves exist and which one is read. That answers it.

The rest, re-checked at 7ff09e6. tsc --noEmit && bun test green: 538 pass, 0 fail. src/client/bots.ts and src/client/plan.ts are byte-identical to c2dea07, so the tables I reproduced last round still stand as measured — base 4161986: 216 rounds · plant 61% · detonated 103 · time 55 · elimination 39 · defused 19 · median 123s · longest 156s; this branch: 223 · 62% · 105 · 21 · 72 · 25 · 123s · 144s. A quarter of rounds decided by nothing becomes a tenth, with the plant rate and the detonation count holding — the improvement is not bought from the terrorists.

Security, for the record. This pull touches four client modules and the README; no route, no request handler, no src/tty.ts, no dependency. I booted this head and put a request through it: / answers 200 in 19 ms carrying the floor src/server.ts declares — default-src 'self', nosniff, strict-origin-when-cross-origin, permissions-policy, session cookie HttpOnly; SameSite=Lax — and a cross-origin POST /api/match is refused 403 before it reads a body. The preview address and the production address both return the platform's 404 from inside my sandbox, production included, so that is my egress and not your branch; the forge's pictures at this head render the page on phone and desktop with the pull's own banner on it.

One judgement I will leave as a judgement, not a request. #105's "the median round comes down" box is still unticked at 123 s, and the pull says so plainly rather than quietly. I do not think anything is owed there: the median is now made of rounds that end in a detonation — a plant plus a 40-second fuse — which is match.ts and explicitly out of that issue's scope, and a round that ends because somebody won it at 123 s is not the same complaint as a round that ends because nobody did. Elimination nearly doubling is the honest cost and it is written down in the README next to the table, where the next person to run the script will find it.

I anchor this on the line that needs the test.

I checked this out at c2dea07 and ran it rather than read it.

What holds up. bun run check is green there — tsc clean, 535 pass. Both measurement tables reproduce exactly on my machine: at the base 4161986, --matches 30 --seed 41 gives 216 rounds · plant 61% · detonated 103 · time 55 · elimination 39 · defused 19 · median 123s · longest 156s; at this head, 223 rounds · plant 62% · detonated 105 · time 21 · elimination 72 · defused 25 · median 123s · longest 144s. 25% of rounds on the clock down to 9%, with the plant rate and detonations holding — the headline is honest and it is repeatable. The new tests are real, not decoration: I mutated the code under them and each one fell where it should — carrier-only pushing breaks the escort test, deleting the hunt goal breaks the defence test, dropping heldInPlace breaks the hold test, and dropping brain.ordered from hunting breaks the existing "a call outranks the rotation" test. plan.ts's time pressure block is the right shape: two numbers, both argued from a measurement, nothing buried three calls deep. No server, route or tty surface is touched; I booted the head locally and / answers 200 in ~6 ms with the CSP, X-Content-Type-Options and Referrer-Policy floor intact, and the screenshots at this head render on phone and desktop.

Fix this first. pushing's if (world.bomb.state === "planted") return false; (src/client/bots.ts, in pushing) has no test: I deleted that line and all 535 tests still pass. It is load-bearing, not belt-and-braces — the old committed could never fire after the plant because it required bot.hasBomb, and widening the rule to the whole side made this line the only thing standing between a terrorist in a post-plant fight and having its feet taken and walked at nearby(bomb.pos, 300) while the policy is mid-duel. That is exactly the "hold near the bomb rather than walk onto it" behaviour chooseGoal is careful about. One test in the shape of underFire does it: plant the bomb, clock low, a terrorist in a fight, expect brain.goal to stay null — the policy keeps the feet. It fails without the guard.

Second, same kind. ${hunting(...) ? ":hunt" : ""} in repath's objective string is unpinned too — remove it and 535 still pass, because in every test the path is empty and stale anyway. In a real round its absence costs up to the 3–6 s repathAt window before the defence leaves its angle, which is the thing the comment above it says it is for. Pin it: a counter-terrorist with a live path and plannedFor set, the clock stepped past HUNT_AT inside its repath window, goal turns to muster.t on the next think. (:called has the same gap today — this is a good moment to close both.)

Small. muster.ct is built, typed and tested but read nowhere in production code; muster.t is the only half anything asks for. Keep it if a use is coming (a terrorist hunting a dropped bomb would want it) and say so in a word, or trim it to what is read.

Not asking for anything here, for the record. The issue's "median round comes down" box is not met — 123 s both sides of the change — and elimination nearly doubles, 39 to 72. Both are stated plainly in the pull with reasons I reproduced and agree with: a detonation is the plant plus a 40 s fuse and match.ts is out of scope, and the idle person is a terrorist in spawn after half time, so a counter-terrorist walking over is the only ending those rounds can have. The one claim I could not verify from the script's output is the "43 of the 55" attribution; nothing rests on it that the tables do not already carry.

This is close. One test for the post-plant guard, one for the hunt's objective string, and I am happy.