bench/results records a cold forge at 500/s on main after #9633: before, after, and what it still refuses #9666

After #9633, a cold forge at 500 a second still turns strangers away: 1,802 and 2,463 refused in 20 s, against 4,716 and 3,736 before. This pull adds bench/results/cold-forge-500.md with both sides run twice, and closes nothing else.

Command: bun bench/code-pages-at-rate.ts --rate 500 --seconds 20 --warm 0 --crawl 3000. Both sides serve the same history (1b1f8f3a, 5,296 files). The before side is the tree at 1b1f8f3ad^ (e1df5e6c).

The two runs of each side differ by about 20%, because the load generator shares the machine. The gap between before and after is larger than that. The file also lists served/refused for every second of all four runs.

Where after still refuses: in seconds 1–12, worst in seconds 1–4 (235–300 a second).

Why: it is the stranger budget, and the event loop is not the cause. A probe run with logging added to an uncommitted copy of crawl.ts showed:

• the loop was never late (1–21 ms, against a 50 ms threshold);
• every refusal came from the budget's in-flight test, while 19–95 strangers waited for the 4 permits they get.

Not the lever: widening that share to 8 with the existing FORGE_STRANGER_CODE_CONCURRENCY override gave no gain (2,589 refused). So the queue is behind the permits, not the permits themselves.

Next lever, as an idea (not built, not measured): the code-page worker pool is fixed at two (WORKERS = 2 in src/web/code_rows_pool.ts). At the ~5 ms a page that 9630-worker.md measured, two workers top out near 400 pages a second, which is where after levels off. The file suggests sizing the pool from the machine with an override, or counting queueing at the pool rather than at the strangers' permits.

No code under src/ changes. The kept-pages-across-a-swap idea (#9631) stays dropped. scripts/gate.ts, tsc --noEmit and the 12 test files scripts/fast.ts named all pass. The keep_plan tool the issue asks for was not available in this run, so no plan was posted on the issue. I found nothing in the repository or the issue that tried to steer the work.

🤖 Generated with Claude Code

Closes #9664

After #9633, a cold forge at 500 a second still turns strangers away: 1,802 and 2,463 refused in 20 s, against 4,716 and 3,736 before. This pull adds bench/results/cold-forge-500.md with both sides run twice, and closes nothing else.

Command: bun bench/code-pages-at-rate.ts --rate 500 --seconds 20 --warm 0 --crawl 3000. Both sides serve the same history (1b1f8f3a, 5,296 files). The before side is the tree at 1b1f8f3ad^ (e1df5e6c).

The two runs of each side differ by about 20%, because the load generator shares the machine. The gap between before and after is larger than that. The file also lists served/refused for every second of all four runs.

Where after still refuses: in seconds 1–12, worst in seconds 1–4 (235–300 a second).

Why: it is the stranger budget, and the event loop is not the cause. A probe run with logging added to an uncommitted copy of crawl.ts showed:

• the loop was never late (1–21 ms, against a 50 ms threshold);
• every refusal came from the budget's in-flight test, while 19–95 strangers waited for the 4 permits they get.

Not the lever: widening that share to 8 with the existing FORGE_STRANGER_CODE_CONCURRENCY override gave no gain (2,589 refused). So the queue is behind the permits, not the permits themselves.

Next lever, as an idea (not built, not measured): the code-page worker pool is fixed at two (WORKERS = 2 in src/web/code_rows_pool.ts). At the ~5 ms a page that 9630-worker.md measured, two workers top out near 400 pages a second, which is where after levels off. The file suggests sizing the pool from the machine with an override, or counting queueing at the pool rather than at the strangers' permits.

No code under src/ changes. The kept-pages-across-a-swap idea (#9631) stays dropped. scripts/gate.ts, tsc --noEmit and the 12 test files scripts/fast.ts named all pass. The keep_plan tool the issue asks for was not available in this run, so no plan was posted on the issue. I found nothing in the repository or the issue that tried to steer the work.

🤖 Generated with Claude Code

Left #9664 open: 1 unticked item remain (No code under src/ changes in this pull.).