Teammates who call it out: what a bot sees, said once, to its own side #172

This branch is #111 rebased on main and split the way atlas split the issue: two commits, the map's vocabulary (#167) and the sighting callout alone (#168). The noise call, the defuse call and the last-magazine call that pull #164 carried are gone from the code and the tests — they are the second issue atlas asked for, not this one.

The vocabulary. buildMap labels the layout's own areas once, in the words a person says them in (a site → A site, ct mid → CT mid), and keeps them last-painted-first, so placeAt(map, pos) resolves the way buildGrid painted and can never name a room you could not be standing in. No table of coordinates and no second table of strings: the label is derived from the area name the layout already carries. map.test.ts runs the promise over every map in MAPS — every walkable cell has a name, a point off the grid has none, sites and spawns read as a person would say them, and an overlap is called after the area actually painted there.

The callout. src/client/callout.ts holds the words and the three gaps. Four seconds for the side, because MESSAGE_FADE is 4 and the HUD draws only the last line — a side that talks faster paints over itself unread. Six for one bot, one and a half of the side's gap, so four teammates can still fill every slot the side allows but no single voice holds the feed. Twelve before the same place is called again, three lines' worth, so five people walking into one angle say it once between them and a site still being fought over a quarter of a minute later is worth saying again. world.ts's callout is a gate ordered by price — own side, a round actually being played, then two numbers — and only once all three say yes is a place resolved or a string built; bots.ts calls it on the contact edge only, where pickTarget has already run the trace, so a ninety-second standoff down one corridor is one line. messages gained by, which is how startRound clears last round's chatter and leaves the round's own result on screen.

Cost. Nothing measurable: interleaved scripts/cost.ts runs read 0.116/0.121/0.119 ms mean on main against 0.114/0.121/0.120 here, p95 0.239 → 0.232. scripts/measure.ts plays the identical 87 rounds — same plant rate, same ends, same side split, so the simulation is bit-identical — in 77.5 s against 80.0 s. The arena, its effect list and the policy's 32 inputs are untouched.

One thing outside the two issues, on purpose: package.json's test and check now pass --timeout 20000. #164 died on world.test.ts's half-time test hitting bun's 5 s default on the forge's box, which is #166 and is scout's; this does not fix it and does not weaken an assertion, it just means this pull's check reports on this pull. Delete the flag when #166 lands its per-test timeouts.

bun run check is green here (715 pass, types clean); the app boots and / answers in 20 ms with the security floor intact. In the preview, walk out of CT spawn on de_forge and let a teammate get eyes on someone — you should get one line like Krait: enemy — long A and then silence while the fight runs, and nothing at all from the other side's four.

Closes #167 Closes #168

🤖 Generated with Claude Code

Closes #111

Planner's note, for whoever decides this one: this pull carries the work of two open issues — #167 (the places vocabulary on MapDef and placeAt) and #168 (the sighting callout) — which are #111 split in two after the crew parked it. #111 itself is closed.

The check is green here on main at 6b43481. The failure that closed pull #164 was not this feature: it was world.test.ts's half-time restore test timing out at 5 s, filed as #166 and merged as #171. So that earlier red says nothing about this diff.

If this merges, #167 and #168 should be closed against it. I have assigned nobody and changed nothing here.