[AUTOMATED] docs(re-pipeline): the quality lane is capped at one builder, and the reason may have expired

[mahaloz]

Round 4's captain measured it while deciding whether to run testers: select -k 3
returns TWO picks against an 11-need backlog, so a builder slot is unfillable at any
depth. 10 of 11 open needs are quality, and every quality need takes the same five
TRACK_RESOURCES leases -- so at most one quality builder ever runs, and two of three
slots idle permanently on the kind of backlog three rounds have actually produced.
Depth is not the constraint; lease structure is.

The serialisation was right when written -- its reason is the silent 85 -> 86
identical-edit merge, which round 1 hit for real (131 vs 129 when the truth was 133).
But two guards have been built since, both at merge time: builder_prompt.md runs
counters --fix INSIDE the merge lease on the rebased tree, re-deriving every count
from the live tree rather than arithmetic; and counters --check runs in the
parity-gates CI job. The merge is already serialised and the numbers already
re-derived, while TRACK_RESOURCES additionally serialises an hour of build to protect
a step that takes minutes and is protected twice.

Written as a CANDIDATE with the experiment that would settle it, not as a change:
two quality builders concurrent, merging serially, counters --check green on main
afterwards. Flags the residual risk as the DIV registry allocation rather than the
counters, since counters --fix re-derives counts and not allocations. Payoff is
roughly 3x on the lane where this backlog lives.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com

🤖 Generated with Claude Code

https://claude.ai/code/session_01YcFmfZndNjgfqLQVBZCdkY