Lateral inhibition: the first claimant suppresses its neighborsSIMULATEDMEASURED

Before, a claim on a file only lowered the stimulus of other intents that write it, so the age term and reinforced thresholds let agents take claimed files anyway and read-set validation sent the later change back. Now ColonyDO.take keeps that soft repulsion and adds winner-take-all: an open intent whose own files carry at least half a fresh claim from other agents is skipped (core chooseIntent, LATERAL_INHIBITION = 0.5). Co-change neighbors still only repel. It is not a lock: there is no lock manager and no release protocol, and a silent agent's claim decays below the threshold after one half-life (about 3.5 minutes; FT_LATERAL_INHIBITION tunes it, off disables it).

• Per-lane landing (as deployed), 1,000 agents: rework 38.7% → 31.2%, landings 85.9 → 86.4/min (+0.6%); locks: 21.7% at 51.9/min. 10,000 agents: rework 38.6% → 31.4%, landings 85.8 → 86.2/min (+0.5%). 100 agents: rework 10.0% → 10.0%, landings 63.4 → 63.4/min (+0.0%): the runs are identical, because with few agents a claimed intent never outranked a free one, so inhibition changed no choice.
• Group commit, 1,000 agents: rework 40.4% → 31.2%, landings 192 → 176/min (-8.4%). 10,000 agents: rework 40.3% → 31.6%, landings 192 → 173/min (-9.8%). Throughput falls because inhibited intents stay at the head of the 200-oldest window until their files' claims clear (95% of the admitted window is inhibited at 1,000 agents), so agents find less independent work. With a 400-intent window (analysis, not built) inhibition costs -2.4% (223 → 218/min), but a wider window raises group-commit rework too (36.3% with inhibition), so the window is a trade-off of its own.
• Draining 1,500 tasks, 1,000 agents: the last task lands sooner (median 2,262 → 1,987 s per-lane, 2,364 → 1,697 s with group commit; locks 2,319 s) and rework falls 35.4% → 25.7% (per-lane); half the tasks land at the same time per-lane (531 → 529 s) and later with group commit (131 → 142 s). No starvation tail appeared.
• Co-change weight under inhibition, 1,000 agents. Weight 5 takes per-lane rework to 26.2% at 86.3/min (weight 2: 31.2%); with group commit it costs throughput (176 → 158/min for 31.2% → 27.6% rework). FT_COCHANGE_WEIGHT stays 2.
• Conflicts vs locks. At 1,000 agents FireTrident with inhibition returns 285 changes per 1,000 submitted against 318 for locks, so the rework gap that is left is the cost per conflict in the model: a returned FireTrident intent is redone from scratch by whoever takes it, a locked change keeps its agent and reworks half. What still comes back is read-set conflicts, which claims on write footprints cannot see: with group commit only 3.1 of 200 returns per 1,000 are write-write (106 before).
• Method. The simulator's take sets each candidate's direct claim to the number of other agents' claims on its write set (live agents heartbeat, so every claim is fresh) and passes LATERAL_INHIBITION to core chooseIntent, as ColonyDO.take does; co-change repulsion uses core neighborRepulsion at weight 2 with hono's last 200 commits. "Window inhibited" is the share of admitted candidates in the 200-oldest window that inhibition held back.