[talwgx]

Engineering ask E90, option B as ruled. Needs log-10x/modules#157 (the declaration of rateReceiverKeepMaxOverTotal); a config key without its declaration fails launch, so modules#157 merges first.

Change

keepMaxOverTotal (option rateReceiverKeepMaxOverTotal, default 500, 0 = unbounded) in both cap-file regulators (rate-object-cap.js, rate-object-lookup-cap.js), modelled on baselineMaxOverTotal:

• A held line of a level in keepCounts whose pattern is under K this window (rg_kc_, unchanged) is a keep line.
• With no service total, or with the option at 0, it passes as before.
• Otherwise it reserves the service total (rg_pas_) like a regular line. If it fits, it passes and counts toward the total.
• If it does not fit, it is past the total and passes only while rg_kx_<container> is under the allowance. The rest take the service's action.

Past the total means what it means for the sample bound: a line the reservation against rg_pas_ refuses. rg_sx_ counts only sample lines the total refused; rg_kx_ counts only keep lines the total refused. Every other path is unchanged: regular lines, the sample bound, K per pattern, and mute-entry lines (the mute path returns before the keep test). With 0, the units added to rg_pas_ match today's for every line.

Worst case per service per node per window with the shipped settings: total + 50 sample lines + 500 keep-count lines (+ mute passes, + warm-up when set). rate-object-local.js and rate-object-lookup.js read no total and are unchanged. config.yaml adds the key and updates the keepCounts and total-file comments. schema.json is regenerated from the modules#157 text the way JsonSchemaGenerator builds it (checked against the untouched baselineMaxOverTotal and baselineCount entries).

Gate: tests/regulator, ghcr.io/log-10x/edge-10x:1.1.150 (sha256:3e755bbc), fix = this branch + modules#157 mounted

The image's bundled rate files and module.yaml are byte-identical to config/modules main, so the main arm is main.

case	main	fix
e1-keep-bound (1,422 new ERROR patterns x 7 lines, cap 1 byte, total 100 lines)	FAIL: 9,954 passed, 0 held	PASS: 650 passed (100 + 50 + 500), 9,304 tier_down
e2-keep-bound-off (option 0)	n/a	PASS: 9,954 passed
e3-keep-bound-mute (e1, mute + cap regulator)	FAIL: 9,954 passed	PASS: 650 passed
k1-errors	PASS 58	PASS 58
k2-mute	PASS 50	PASS 50
k3-mute-nocap	PASS 50	PASS 50
k4-keepcounts-unset	n/a	PASS 58
u1-events	PASS 40	PASS 40
u2-unit-typo	PASS (launch fails)	PASS
u3-unit-typo-mute	PASS (launch fails)	PASS
m1-mute-counts	PASS (others 50 of 80)	PASS (others 50 of 80)
t3-reservation	PASS 19,939 / 20,000 B	PASS 19,939 / 20,000 B
g0-gitops-off	PASS 300	PASS 300
t1-total	PASS 247 (51,422 B)	PASS 247 (51,422 B)
t2-noisiest	PASS 346	PASS 346
s1-star	PASS 300	PASS 300
d1-defaults	PASS 66	PASS 66
b1-sample-bound	PASS 100	PASS 100
b2-sample-bound-off	n/a	PASS 400
b3-sample-bound-mute	PASS 100	PASS 100
w1-windows	PASS 192 / 213	PASS 192 / 213
w2-windows-mute (windows 1, 2; total 517)	PASS 547, 517	PASS 547, 517
w3-windows-scaled (windows 1, 2; total 301)	PASS 312, 301	PASS 312, 301
w4-windows-late (windows 1 to 4; total 301)	PASS 21, 309, 301, 301	PASS 21, 309, 301, 301

Every existing case passes on the fix with the same counts as main. The paced cases w2 to w4 ran after the rest, with run.sh's wait for the Forward input's listening line, which needed no change.

Harness

• Three cases e1 to e3.
• run.sh runs each engine with docker run -d --name $PREFIX-... and removes it with docker rm -f -v (it no longer uses docker run --rm in the foreground).
• A paced case waits for the engine's "Forward protocol server listening" line before sending.
• The default image is now 1.1.150.

🤖 Generated with Claude Code