Skip to content

Ranked readers move warm pulls for price and settle - #8986

Merged
kixelated merged 8 commits into
mainfrom
quest/m3/cache-routing/settle
Nov 17, 2026
Merged

kixelated merged 8 commits into
mainfrom
quest/m3/cache-routing/settle

Conversation

@kixelated

@kixelated kixelated commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator

Finishes quest/m3/cache-routing/settle.md and deletes it.

Result

The quest's first candidate holds: ranked readers move a pull to a cheaper warm copy that turns up later, and settle. TLC checks it, and the same check fails for the price-moving pick and for the ranked rule with its rank removed. Settled pulls, the first cut select.md named, pass the same check.

Ranked readers cut the first cut's drift over the cheapest tree by about two thirds on drones and random meshes. They also merge simultaneous starts, and let a drone that loses the uplink hand it to one drone instead of four. The price is a rank on three messages plus one update message, and about 2.6 times the warm messages between peers. In the simulator, unfed reader-hops rise 15% to 25% under churn, but that comes from the simulator's clock, not from the moves (see Churn). The maintainer picked ranked readers on 2026-11-17, and select.md is now the plan to port them.

No production code and no wire change: the rule lives in the warmth simulator (rs/moq-net/src/route/tests/warm.rs), and the model in rs/moq-net/verify/warm/.

The rule

  • Rank. A relay with a reader of its own for an Instance ranks it by its route cost to the Instance, then its Node ID.
  • What travels. A pull carries the lowest rank among the readers it feeds (its relay's own and each downstream pull's), with an update when that changes. An advert and a subscribe answer carry the lowest rank their sender feeds besides their receiver.
  • Choice. A relay takes a warm copy only from a neighbour feeding a reader ranked strictly below every reader the relay feeds. Such a copy is fed whatever the relay does, so it costs the link's cold price alone. The relay heads for the cheapest such copy or its route, make before break, whenever that set or its own rank changes. No price above the first hop enters the choice, and there is no hysteresis.
  • Why it settles. A relay's choice reads only copies feeding lower ranks, and what those feed never depends on a higher rank's choice. Once outside events stop, the lowest rank settles, then the next, by induction. A relay's own demand never comes back to it as a price, which is the reflection every fence of the pick chased.
  • What stays from the first cut. Adverts, the mark, the path in every offer, the offer on the subscribe answer, and choosing again on a route change. The advert audience is still priced from the publisher: offer plus the link's cold price, against the neighbour's route.
  • A subscription with no rank was not measured: no sweep or model here has a relay that sends none. In the simulator such a pull is dropped from a relay's rank. select.md gives the port one rule to build instead (an endpoint's is a reader of the relay it reaches; an older relay's pins that relay to its route) and requires the mix in the port's sweep and model.

Candidates tried

candidate outcome case
ranked by route cost, copy at the link alone, audience priced from the publisher holds: 0 of 6,000 churned meshes unsettled, TLC passes this PR's advertise_not_down+hops+order_route
the same, audience priced at the link alone settles, but warm messages grow with a full core mesh: 0.9 / 4.4 / 12.3 / 28.6 per reader change at 4 / 8 / 16 / 32 cores, and a crowd on 64 peered drones costs 173 against 63 useful_adverts_do_not_grow_with_the_core_mesh now fails on it (28 against 124)
the same, a copy priced at the parent's summed offer plus the link, kept while no dearer (+summed) settles, but reads demand above the first hop, so relays keep dearer copies: drones +18.4% over the cheapest against +4.9% ORDER_SUMMED in the churn sweep
ranked by Node ID alone (+order_hash) settles (Hash.cfg; 0 of 6,000 unsettled under the earlier link-priced audience), but a near pop never hears a middle pop ranked above it, so pops keeps the path join's tree a_join_skips_pops_without_demand_and_warm_ones_draw_it
ranked, no path in offers (advertise_not_down+order_route) settles on the same trees, but a cycle formed on stale adverts lasts longer: 99 deliveries on drones against 29 churn at 2,000 per kind
ranked with the rank check removed never settles: readers at two drones each take the other's warm copy over their uplinks, find their own path in what they now pull, drop it, and repeat Orderless.cfg, an open finding
the price-moving pick, rejected in #8903 never settles case 503 Pick.cfg, an open finding

The quest's other candidates were not built. Pricing a copy at what it adds to the tree needs an exact price and non-overlapping moves, which #8903's sole-demand list already failed. Counting only relays with their own reader as warm gives up coalescing at every CDN core. Ranked readers settled without either.

The quest feared that a relay that only forwards would move the higher-ranked pulls it feeds when a lower rank arrives. It does move them, but always toward a strictly lower rank, which is what the induction needs. TLC found no loop from it.

The two cases and the quest's shapes

settled pulls, first cut ranked readers the price-moving pick path joins cheapest
pops, five phases summed 71 59 61 69 59
drones in turn 75, one uplink 75, one uplink 75, one uplink 155, five 75
drones at once 51, five uplinks 19, one 51 51 19
drones one / two hops apart 41 / 33, at 69 / 62 ms 17 / 17, at 81 / 69 ms 33 / 25, at 81 / 69 ms 41 / 41, at 69 / 62 ms 17 / 17
drones lose the uplink 118, four uplinks, a 17-step cycle 94, one uplink, no cycle 118, four uplinks 198 94
edges in turn / at once 13 / 9 13 / 8 12 / 9 14 / 9 12 / 7
one region with unequal uplinks 29 29 24 29 24
ring in turn / at once 2904 / 392 2904 / 112 2904 / 392 9624 / 392 2904 / 112
tiers, a crowd / one viewer 1095 / 1096 1095 / 1096 1095 / 1096 1191 / 1096 1095 / 1096
warm messages per reader change, pops / drones at once / ring at once / tiers crowd 1.20 / 4.00 / 2.00 / 3.38 2.20 / 7.20 / 6.94 / 5.33 0

"The price-moving pick" is #8903's advertise_not_down_query+hops, which also queries, hence its 24 on the region with unequal uplinks. Every first frame matches the path join's except drones one hop apart, 81 ms against 69, where the second drone first pulls its uplink and then moves onto the first drone.

Churn

400 churned meshes per kind, random / drones / regions:

settled pulls, first cut ranked readers the price-moving pick path joins
settled trees over the cheapest +3.4% / +16.2% / +5.4% +1.2% / +4.9% / +4.4% +2.7% / +9.4% / +4.9% +2.7% / +28.9% / +5.0%
unsettled of 2,000 per kind 0 / 0 / 0 0 / 0 / 0 1 / 0 / 0 (random 503) 0 / 0 / 0
settled trees over the cheapest, 2,000 per kind +3.2% / +16.0% / +5.3% +1.4% / +6.3% / +4.6%
longest cycle, deliveries 20 / 18 / 14 19 / 23 / 18 (35 / 29 / 26 at 2,000) 18 / 39 / 24 16 / 14 / 17
unfed reader-hops 19,008 / 11,691 / 17,470 21,917 / 14,584 / 20,661 18,000 / 12,381 / 22,472 15,700 / 7,725 / 15,735
warm messages 9,020 / 24,914 / 4,655 16,980 / 33,950 / 13,133 9,909 / 24,937 / 21,151 0
of which rank updates 0 7,876 / 9,497 / 8,494 0 0
messages 65,409 / 82,961 / 61,602

The extra unfed reader-hops are not readers stalled by moves, which are make before break. Classifying each unfed reader-round by the first broken hop of its chain (a temporary diagnostic, not committed):

  • Readers already fed lose nothing. Their unfed hops are equal or fewer under ranked readers: 5,388 against 5,443 on random meshes, 1,977 against 2,098 on drones.
  • The whole rise is joins waiting for a first frame: 18,307 against 15,033 on random meshes, and 13,477 against 10,273 on drones. Most of it is the reader's own subscribe or answer still in flight.
  • The cause is the simulator's clock. It delivers one control message per step across the whole mesh, so every rank and advert message anywhere delays every join. Unfed hops per delivery match across arms: 0.319 against 0.317 on random meshes, 0.174 against 0.160 on drones, and 0.304 against 0.316 on regions. Path joins sit at 0.302, 0.155, and 0.319.
  • Not re-aimed pending pulls. Keeping a pending pull while its target stayed an option changed nothing (21,809 against 21,917 on random meshes).

select.md now requires the port's acceptance tests to measure unfed reader-hops and first frames with each link delivering at its own pace, beside path joins, before this figure is read as production behaviour.

Growth per link kind

Warm messages per reader change, from benchmark_warmth_growth:

link kind settled pulls, first cut ranked readers the price-moving pick
unpeered CDN, one viewer, 4 to 32 cores, 1 to 4,096 broadcasts 0 0 1.0
unpeered CDN, a crowd, 4 to 32 cores 0 0.67, flat 1.0
peered edges, a crowd on 16 broadcasts, 3 / 16 / 64 edges per core 2.0 / 15.0 / 63.0 4.0 / 40.0 to 42.5 / 164.8 4.2 / 19.3 / 77.9
peered drones, a crowd, 5 / 16 / 64 drones 4.0 / 15.0 / 63.0 7.1 / 39.2 / 161.1 4.8 / 16.7 / 70.0

Between peers every relay that could take the copy hears each rank change, so ranked readers cost about 2.6 times the first cut there, still linear in the peers. Trees and first frames match the first cut in all 71 growth rows; the busiest relay holds as much or, in four peered rows, less.

Model check

rs/moq-net/verify/warm/WarmSettle.tla, with the verify-warm skill. It models where each relay pulls one broadcast from. Whatever one relay tells a neighbour over a session (its pull's rank, its answer, its advert) is a register the neighbour reads late, so relays act on stale words for as long as TLC likes. Readers start and stop, and uplinks close and reopen, up to a bound; then nothing outside changes. With weak fairness on every delivery and reaction, it checks PullsQuiet (<>[][UNCHANGED <<up, nxt>>]_vars: judged from when the outside events stop, not from when queues drain) and EventuallyFed (every reader with a route ends fed through live answers, so two relays waiting on each other fails it).

config rule result
Drones.cfg ranked passes, 158,198 states, three drones, three reader changes
Hash.cfg ranked by Node ID passes, 158,198 states
Flap.cfg ranked passes, 44,427 states, readers at drones 2 and 3 while drone 2's uplink closes and reopens, changing its rank
Mesh503.cfg ranked passes, 490,125 states, churned random mesh 503's six relays and five readers from no pulls
Settled.cfg settled pulls passes, 407,526 states
SettledDrones.cfg settled pulls passes, 6,852 states
Pick.cfg the pick fails PullsQuiet within 6 minutes (open finding)
Orderless.cfg ranked, no rank check fails PullsQuiet within 9 minutes (open finding)

The configs take about an hour together, so they get their own nightly tla leg (warm) rather than joining rest. Two open findings keep the check honest: if either starts passing, nightly fails.

Settled pulls passing closes the open question select.md had: the model found no warm retries that repeat, so the first cut takes no fallback.

Wire and relay state over the first cut

  • Wire, all additive, for the then-current lite wip: a rank on each SUBSCRIBE and a message updating it; a rank on each advert and on the subscribe answer and its updates. A rank is a route cost and a Node ID. Nothing else changes: adverts keep the offer cost (the audience test needs it), the path, and the mark.
  • Relay state: one rank per served subscription, one per heard advert or answer, and the rank last told on each pull. All of it is bounded as the adverts are: per neighbour per Instance the relay has a route for.
  • A subscription with no rank was not measured: no sweep or model here has a relay that sends none. In the simulator such a pull is dropped from a relay's rank. select.md gives the port one rule to build instead (an endpoint's is a reader of the relay it reaches; an older relay's pins that relay to its route) and requires the mix in the port's sweep and model.

Decisions

  • Maintainer, 2026-11-17 (comment): ✅ the port uses ranked readers (rejected: settled pulls, which give up about two thirds of the drift for a small wire saving). This was this PR's ## Waiting item, recommended here.
  • Maintainer, 2026-11-17 (same comment): ✅ warmth ships in the next wip version, after the lite-07 cut (rejected: holding lite-07 for the port; freezing the rank and advert fields in lite-07 before the port has exercised them). The 2026-11-14 decision stands: lite-07 freezes without a warm wire.
  • Rewrote select.md as the plan to port ranked readers, with settled pulls as the measured, rejected alternative, both decisions in its Plan, and quest/m2/lite07-finalize.md added to its Required, since the port is the first item after the cut. It also carries two requirements the spike left open. One: get the model bound that did not finish (a reader change during an uplink flap) to finish, by a smaller mesh or a symmetry reduction, or say why not and what covers it. Two: measure unfed reader-hops with per-link delivery. The epic README and the verify-warm skill name the rule too.
  • Built the coordinator's first candidate, then attacked it as the quest asked. It held; the variants above failed or cost more.
  • Ranks by route cost then Node ID (the quest's first key) rather than by Node ID alone: both settle, but the hash order loses the pops case.
  • Prices a copy at the link's cold price alone and keeps no hysteresis: summed pricing reads demand above the first hop and drifts three times as far on drones.
  • Prices the advert audience from the publisher, unlike the choice: at the link alone a warm core told every core it meshes with, growing linearly in cores.
  • Keeps the path in offers: without it the rule still settles, but stale-advert cycles last three times longer on drones.
  • Keeps the rejected and failing rules as measured arms (ORDER_HASH, ORDER_PATHLESS, ORDER_SUMMED) and as open-finding configs, so the tables can be rerun.
  • Moved the model check select.md had asked of the port into this spike: select's Plan now ports whichever rule verify-warm checks, and moves its Coverage table onto the port.
  • Gave the warm configs their own nightly leg rather than adding about an hour to rest.
  • Matched the model's tie-break between equal sources to the simulator's (the highest rendezvous hash, relay IDs standing in), from review round 1; every config kept its result.
  • Removed PickDrones.cfg from the spike's first commit: the pick settles on three drones, so that config proved nothing. Pick.cfg now runs it on case 503, where it fails.

Impact

Public API: none. Wire: none in this PR, which is test-only. It touches the warmth simulator (rs/moq-net/src/route/tests/warm.rs, warmth.rs), adds the TLA+ model's configs and skill, and adds one nightly leg. The wire above lands with select.md's port in the first wip version after the lite-07 cut; nothing lite-07 freezes changes.

Validation

  • cargo test -p moq-net --lib route::tests::warmth: 14 passed. New tests: a_ranked_pull_follows_a_pop_that_turns_warm_later, ranked_drones_starting_at_once_keep_one_uplink, ranked_pulls_settle_random_case_503, and ranked readers in useful_adverts_do_not_grow_with_the_core_mesh. That last assertion fails with the link-priced audience (28 against 124 warm messages), checked by reverting it.
  • benchmark_warmth_cases, benchmark_warmth_churn (400 per kind, and 2,000 per kind for the ranked arms), benchmark_warmth_growth: the tables above.
  • TLC on every config under rs/moq-net/verify/warm/: the table above. A wider Flap (a reader change during the flap, or two reader changes and two flaps) passed 2M to 3M states without finishing in over an hour, so it is left out.
  • test/route-sweep.sh's warmth line took 23 minutes on a host at load 31, 17 of them in the growth sweep with all four ranked arms in it; it now runs only the main ranked arm. The script's "about 10 minutes" predates this PR and was not remeasured on an idle host.
  • just check passes; quest check --base origin/main passes.
  • Merged origin/main after feat(moq-net): the multicast route rule as a sans-IO router #8901 landed: no conflicts, warmth tests pass.

Follow-ups

  • None new. select.md owns the port of ranked readers, Required on the lite-07 cut, with no open decision.

Review fixes by the coordinator, 2026-11-17

  • A subscription with no rank (Pullfrog, and the automated OpenAI review): the plan and the simulator described different rules and neither was exercised. select.md now states the rule to build and requires an unranked relay in the port's churn sweep and model.
  • The model's delivery (the automated OpenAI review): one register per neighbour delivers a neighbour's rank, answer, and advert together, so select.md no longer says "every word late and in any order" and asks the port to cover independent delivery.
  • The skill's sync commit is the PR's base, as REVIEW.md asks.

Models

Size: L
Implementer: Claude Code, claude-opus-5-5, unknown effort
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, M)
Implementer: Claude Code, claude-fable-5-2, high effort (review fixes)

Review

Rewritten by quest review after each round.

  • Head: dd31e33

  • Base: origin/main at 036b978

  • Risk: high, it changes .github/workflows/nightly.yml, which is always high: workflows run with the repository's secrets and decide what merges and ships

  • Classified: high by the always-high paths

  • Reviewer: Codex, gpt-6.1-sol, xhigh effort

  • Session: 01a1e959-b2b6-7a30-bdfb-fcf0a464ca85, round 4

  • Score: 10/10, The incremental documentation changes accurately clarify compatibility and model limits, and the new sync commit preserves the covered simulator baseline; no builds or tests were run.

  • Rounds: 4; first round 2 (medium 1, low 1); findings 2 (medium 1, low 1): fixed 2, refuted 0, open 0

  • Findings: none

kixelated and others added 5 commits November 16, 2026 21:18
…g model

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lisher

Advert audiences under ranked readers are priced from the publisher, so a
full core mesh's warm messages stay flat. WarmSettle gains configs for
ranked readers, settled pulls, and two open findings: the pick, and the
ranked rule without its rank. select.md names ranked readers beside settled
pulls; the settle quest is done.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…growth

Flap holds two readers fixed while drone 2's uplink closes and reopens:
44,427 states, where a reader change during the flap passed 2M states
without finishing. The growth sweep keeps ranked readers' variants out,
since they change trees rather than how adverts grow.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@kixelated kixelated changed the title Warm pulls move for price and still settle [L] Ranked readers move warm pulls for price and settle Nov 17, 2026
@kixelated kixelated changed the title [L] Ranked readers move warm pulls for price and settle Ranked readers move warm pulls for price and settle Nov 17, 2026
The simulator takes the neighbour with the highest rendezvous hash among
equal sources; the model took the lowest relay. Relay IDs stand in for
those hashes, so the model now takes the highest. Every config keeps its
result. The skill's sync commit is the PR's base.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@kixelated

Copy link
Copy Markdown
Collaborator Author

Maintainer decision (2026-11-17, asked in the coordinating session with this PR's results):

  1. The port uses ranked readers, not settled pulls. Picking it also picks this PR to merge.
  2. Warmth ships in the next wip version, after the lite-07 cut. The 2026-11-14 decision stands: lite-07 freezes without a warm wire. The cut waits on the switch, and the port in select.md is the first item after it.

Rejected: settled pulls (gives up about two thirds of the drift for a small wire saving); holding lite-07 for the port; freezing the rank and advert fields in lite-07 before the port has exercised them.

(Written by Claude Fable 5.2 with high effort)

The maintainer picked ranked readers over settled pulls on 2026-11-17, and
warmth's wire for the first wip version after the lite-07 cut. select.md
becomes the plan to port the one rule, now Required on the cut, and carries
two requirements the spike left open: the model bound that did not finish,
and measuring with per-link delivery the unfed reader-hops the simulator's
one-message-per-step clock inflates.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@kixelated
kixelated marked this pull request as ready for review November 17, 2026 12:57
@kixelated

Copy link
Copy Markdown
Collaborator Author

Applied in bedbbd1, replying to the decision:

  • quest/m3/cache-routing/select.md is now the plan to port ranked readers: title, Goal, Plan, wire, and relay state name the one rule, with settled pulls as the measured, rejected alternative. Both decisions are recorded in its Plan, dated 2026-11-17, and its ## Waiting item is gone. The wire lands in the first wip version after the cut, and quest/m2/lite07-finalize.md joins its Required.
  • Its Plan requires the port to get the unfinished model bound (a reader change during an uplink flap) to finish, by a smaller mesh or a symmetry reduction, or say why not and what covers it.
  • It also requires measuring unfed reader-hops with per-link delivery. The extra unfed hops turned out to be the simulator's clock, not moves. Readers already fed lose nothing (5,388 against 5,443 on random meshes); the rise is joins waiting behind the added rank and advert messages, since the simulator delivers one message per step across the mesh. Unfed hops per delivery match across arms (0.319 against 0.317).
  • The epic README and the verify-warm skill name ranked readers too. No production code, no wire.

(Written by Claude Opus 5.5)

@coderabbitai

coderabbitai Bot commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 17 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used all 4 included reviews currently available.

Learn how review limits work.

Review configuration:

Run configuration

  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: Organization UI
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: b7eea0d3-b8d2-438c-a56c-ab47e5ca1714
Commits

Reviewing files that changed from the base of the PR and between 9143829 and dd31e33.

Files selected for processing: 18

  • .claude/skills/verify-warm/SKILL.md
  • .github/workflows/nightly.yml
  • quest/m3/cache-routing/README.md
  • quest/m3/cache-routing/select.md
  • quest/m3/cache-routing/settle.md
  • rs/moq-net/src/route/tests/warm.rs
  • rs/moq-net/src/route/tests/warmth.rs
  • rs/moq-net/verify/README.md
  • rs/moq-net/verify/warm/Drones.cfg
  • rs/moq-net/verify/warm/Flap.cfg
  • rs/moq-net/verify/warm/Hash.cfg
  • rs/moq-net/verify/warm/Mesh503.cfg
  • rs/moq-net/verify/warm/Orderless.cfg
  • rs/moq-net/verify/warm/Pick.cfg
  • rs/moq-net/verify/warm/Settled.cfg
  • rs/moq-net/verify/warm/SettledDrones.cfg
  • rs/moq-net/verify/warm/WarmSettle.tla
  • test/route-sweep.sh
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

@kixelated kixelated left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Automated review by review (OpenAI)
Reviewed commit: bedbbd1e4fc29a70eca9d8c50be50aea385feb3c

  1. quest/m3/cache-routing/select.md:58–61: narrow the model claim and carry the missing case into the port. TLC delivers one neighbor’s rank, answer, and advert together as its newest register. It can delay neighbors independently, but it cannot see a new rank alongside an old answer or advert from that same neighbor. Those ride separate streams in the proposed protocol. Replace “every word late and in any order” with the actual scope, and add a port acceptance case for independent delivery (including a stale advert during rank withdrawal or replacement).
  2. rs/moq-net/src/route/tests/warm.rs:783–789: cover ranked and legacy demand on the same relay. The plan says an unranked downstream pull wins the rank comparison, but the simulator’s minimum discards None pulls. With both a legacy pull and a ranked reader, this helper would choose the ranked reader’s rank and could allow a warm move. Treat legacy demand separately from “no reader” in the port, and add a mixed-version test. The current homogeneous sweeps don’t exercise this.

The ranked rule looks worth porting: the spike compares it with credible alternatives and keeps failing controls. Neither gap blocks a test-only spike if carried forward explicitly. Static review; I did not rerun TLC or the sweeps.

@pullfrog pullfrog Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ℹ️ No critical issues. Two doc-level points are inline.

Reviewed changes

This review covers the whole spike: the ranked-readers rule in the warmth simulator, the new WarmSettle.tla model with its configs, skill, and nightly leg, and the rewrite of select.md as the plan for the port.

  • Ranked readers in the simulator: Arm::order and Arm::summed add a rank to subscribes, answers, and adverts, plus a Msg::Rank update. A relay takes a warm copy only when the copy's rank is strictly below the relay's own, and it prices that copy at the link alone.
  • New arms and tests: ORDER_ROUTE, ORDER_HASH, ORDER_PATHLESS, and ORDER_SUMMED join the churn gate and the benchmarks; the growth sweep keeps only ORDER_ROUTE. Three new tests cover pops, drones starting at once, and case 503.
  • TLA+ model: WarmSettle.tla checks PullsQuiet and EventuallyFed under late, stale register delivery. Pick.cfg and Orderless.cfg are open findings that must keep failing, and the configs get a dedicated warm leg in the nightly tla matrix.
  • Plan rewrite: select.md now plans the port of ranked readers after the lite-07 cut, settle.md is deleted, and the epic README records the decision.

I checked the model against the simulator line by line, and the following match: Offer (the warm link under the rank), Heard (the parent's answer when it sends no advert), the rank filter in Options, the Better tie-break (relay IDs stand in for the rendezvous hash), and the arms of Decide against ordered. cargo test -p moq-net --lib route::tests::warmth passes at the head (14 tests, 4.8 s), and so does sh/coverage-tables.py. TLC could not run in this sandbox, so the model's results are taken from the PR body.

Pullfrog  | Fix all ➔ | Fix 👍s ➔ | View workflow run | Using claude-opus-5-5 | 𝕏

Comment thread .claude/skills/verify-warm/SKILL.md Resolved
Comment thread quest/m3/cache-routing/select.md Resolved
@moq-bot
moq-bot Bot disabled auto-merge November 17, 2026 13:04
@moq-bot

moq-bot Bot commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

quest land watch: ended this pull request's pick, disarming its auto-merge, since its head bedbbd1 is held by finding by pullfrog[bot]: The sync commit 9143829 is the branch's fork point from before it merged origin/main. The PR's base is now 1399add (#8986 (comment)); finding by pullfrog[bot]: This line and the simulator describe different rules, and the port will be built from this line. Everywhere else in this (#8986 (comment)). Push a fix, or answer each by a thread reply or a comment linking its review or decision, then land it again.

(written by quest land watch)

The plan said a peer with no rank ranks below every reader, while the
simulator drops such a pull from a relay's rank, and neither the sweep nor
the model ever had an unranked relay in a mesh. The plan now states one rule
for the port to build and requires the mix in its sweep and model. It also
narrows what the model's one register per neighbour covers and asks the port
for independent delivery. The skill's sync commit moves to the PR's base.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@kixelated

Copy link
Copy Markdown
Collaborator Author

Answering the automated review of bedbbd1e4f (#8986 (review)), both fixed in dd31e33:

  1. The model's scope. select.md no longer says "every word late and in any order". It says what the model delivers: one register per neighbour, so a neighbour's rank, answer, and advert arrive together, any number of steps stale, with neighbours falling behind independently. The port's requirements gain independent delivery: a new rank before or after its advert and answer, a stale advert outliving a rank's withdrawal or replacement included, by splitting the register in the model or pinning the cases in acceptance tests with each stream at its own pace.
  2. Ranked and legacy demand on one relay. Same finding as Pullfrog's thread on select.md. The plan now states the rule the port builds (an older relay's pull pins the relay it reaches to its route), says the simulator's rank helper does not build it, and requires the mixed-version case in the port's sweep and model.

Also answering the maintainer decision in #8986 (comment): the head applies it. select.md plans the port of ranked readers for the first wip version after the lite-07 cut.

(Written by Claude Fable 5.2 with high effort)

@kixelated kixelated left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Automated review by review (OpenAI)
Reviewed commit: dd31e33d6c9e195ff866ba9015e891c777f9b4b2

Both findings are addressed for this spike. The plan now states the model’s atomic-delivery limit and requires a stale-advert/independent-delivery case in the port. It distinguishes an endpoint, which the relay ranks locally, from a legacy relay, which pins the upstream choice to the route, and requires mixed demand in both the churn sweep and model. These are explicit port requirements, not evidence the existing model already covers them. No new blocker in this delta. Static review only.

@pullfrog pullfrog Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ No new issues found.

Reviewed changes

This review covers the docs-only delta since the prior Pullfrog review (bedbbd1..dd31e33).

  • Fixed the skill's sync commit: verify-warm/SKILL.md now names the PR base 1399add3c8, as REVIEW.md asks.
  • Stated the rule for a subscription with no rank: select.md makes an endpoint's pull a reader of the relay it reaches and pins a relay behind an older relay to its route. It also says the simulator's WarmSim::rank drops such pulls and does not build this rule.
  • Added port requirements: the churn sweep and model must include an unranked relay beside a ranked reader, and independent delivery of rank, advert, and answer must be covered.
  • Narrowed the model claim: select.md now says the model delivers each neighbour's words together as one stale register, not "every word late and in any order".

Both open Pullfrog threads were already resolved by the author and match the new text.

Pullfrog  | View workflow run | Using claude-sonnet-5-5 | 𝕏

@kixelated
kixelated added this pull request to the merge queue Nov 17, 2026
Merged via the queue into main with commit aa780b3 Nov 17, 2026
10 checks passed
@kixelated
kixelated deleted the quest/m3/cache-routing/settle branch November 17, 2026 13:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant