b0cf7a1 · hyperswarm-peer

hyperswarm-peer commits / b0cf7a1

docs: price the rest of defect B — both halves, from the reference's constants

hyper-games #73 NEXT item 7 asked for the three-generation _serverRelays cycling
and refresh-chain.js's reverse hash chain. Reading persistent.js and index.js
prices both, and neither is a correctness defect.

§9.2 left an explicit open question: does an empty `refresh` make a node treat a
re-announce as a new record? It does not. onannounce validates the signature and
calls records.add identically either way; `refresh` only registers a token so a
LATER renewal can arrive as a tiny unsigned announce that _onrefresh expands
from the cached record. Sending it empty costs a full signed announce per
renewal instead of about forty bytes — bandwidth, not behaviour.

⛔ And the chain carries a failure mode the full announce does not: it is a
fixed number of blocks built backwards from a seed, and a peer that exhausts it
without falling back stops renewing SILENTLY. Taking on a silent-stall risk to
save bandwidth is the wrong direction to trade in.

Generation cycling costs at most one record TTL, and the reference states it:
defaultMaxAge = 20 minutes (index.js:614). Our refresh cadences are 10 min ± 2
and 5 min, both inside it, so a holder we stop refreshing carries a REDUNDANT
copy for up to twenty minutes — an extra copy, not a wrong answer. The case
where it would be a wrong answer is the peer going away, and that is the
liveness ping, which is built.

Same discipline as A2 an hour earlier: the reading says what it would cost to
build, and the cost says do not build it yet. Priced, not guessed, and not
claimed to be invisible on the public DHT at scale.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

b0cf7a1385e29f8faf6589b06264d7e5055d3fc9

by plankatron <fleeky@gmail.com> · 2026-09-10T19:53:54.000Z

Changed files (1) +41 -4

docs/PARITY-hyperswarm.md +41 -4

…	…	@@ -925,10 +925,47 @@ up to `SELF_ANNOUNCE_UPDATE_MS` instead of ~3 seconds.
925	925	- a ping round on a 3s schedule with tag→response accounting
926	926	- re-announce immediately when answered < min(held, 3)
927	927
928		⏳ Also still absent: the three-generation `_serverRelays` cycling and
929		`refresh-chain.js`'s reverse hash chain (we always send `refresh` empty).
930		Neither is required for the record to stay alive; both are needed for full
931		parity.
	928	### 12.2 The rest of B, priced by reading (2026-09-10)
	929
	930	The two remaining pieces were carried as open parity gaps. Reading
	931	`persistent.js` and `index.js` prices both, and neither is a correctness defect.
	932
	933	The refresh chain is an OPTIMISATION, and §9.2's open question is closed.
	934	§9.2 asked whether an empty `refresh` makes a node treat a re-announce as a new
	935	record. It does not. `onannounce` (persistent.js:100-150) validates the
	936	signature and calls `this.records.add(k, publicKey, record)` on every full
	937	announce, identically, whether or not `refresh` is present. All `refresh` does
	938	is `this.refreshes.set(hash(refresh), {k, record, announceSelf})`, registering a
	939	token so a LATER renewal can arrive as a tiny announce with no peer and no
	940	signature at all — `_onrefresh` (`:72-97`) then re-adds the cached record and
	941	rotates the token.
	942
	943	So sending it empty costs a full signed announce per renewal — peer record plus
	944	a 64-byte signature, and an Ed25519 verify at each of ~20 holders — where JS
	945	pays about forty bytes and a hash lookup. Real, and bandwidth rather than
	946	behaviour.
	947
	948	⛔ It also carries a failure mode the thing it replaces does not: the chain is a
	949	fixed number of blocks built backwards from a seed, and a peer that exhausts it
	950	without falling back to a full announce stops renewing SILENTLY. That is a
	951	correctness risk taken on to save bandwidth, which is the wrong direction to
	952	trade in — so it is priced here and not built.
	953
	954	Generation cycling costs at most one record TTL. JS explicitly unannounces
	955	at any node that took generation 1's announce but not generation 2's
	956	(announcer.js:192-196). We do not, so such a node keeps our record until it
	957	expires on its own. That bound is stated in the reference:
	958	`defaultMaxAge = 20 * 60 * 1000` (index.js:614) — 20 minutes.
	959
	960	Our cadences sit inside it: `TOPIC_REFRESH_INTERVAL_MS` 10 min ± 2 min jitter,
	961	`SELF_ANNOUNCE_UPDATE_MS` 5 min. So a record never lapses while we want it, and
	962	a holder we stop refreshing carries a redundant copy for at most 20 minutes.
	963	That is an extra copy, not a wrong answer — a lookup through it still returns a
	964	peer that is there. The case where it WOULD be a wrong answer is the peer going
	965	away, and that is the liveness ping (§12), which is built.
	966
	967	⚠ Not claimed: that either would be invisible on the public DHT at scale. What
	968	is claimed is the price, from the reference's own constants.
932	969
933	970	## 13. Slice 9 — defect D fixed: associate the topic before the early returns
934	971