The panel body listed the address and preimage, but its heading still read
`0x134e73f0…df59` — the one identifier a miner does not recognise, given top
billing. It now falls back through name, address, preimage, which is the order
a reader can actually use them in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
A reward preimage is exact and unrecognisable — no miner knows theirs by sight.
The address is the string they configured and the one their balance shows
against, and the two are related by a derivation the chain performs anyway:
`pallets/mining-rewards` pays `qp_wormhole::derive_wormhole_address`, which is
`qp_poseidon_core::rehash_to_bytes`. Running the same function reproduces the
address exactly, with no RPC and no lookup table.
The derivation takes the preimage and nothing else — no genesis, no chain
parameters — so one preimage is one account on every Quantus chain. The unit
test pins it against a real pair (this observer's own node and the address it
was paid at), so a dependency upgrade that changed the derivation fails the
build rather than quietly relabelling every miner on the site.
Unnamed miners now render as their address. The miner panel shows all three
identifiers, because they answer different questions: the node name is an
inference and may be absent, the address is what the miner recognises, and the
preimage is what the header actually carries and the only one the chain
asserts.
SS58 is written out rather than pulled from `sp-core`, which would bring a
substrate runtime's worth of dependencies to format 36 bytes — the same trade
the readme records for declining subxt. Base58 is hand-rolled because it is
long division and nothing else; blake2 is not, because checksums are not worth
hand-rolling. Both are covered by the address test vector.
`RecentBlock` gains the attribution source: the ticker cannot tell a node name
from an address by looking, since `quanpool-payout-mainnet` and `quantus-radar`
both begin with the letter Quantus addresses do. Two tests that asserted the
old display policy now assert the new one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Live telemetry names the miners who win often and leaves everyone else as a
hex string. Measured on mainnet: of 22 miners, 2 were named and 14 had no name
available anywhere — but 6 already had one on Planck, under the same reward
preimage. Those 6 are exactly the small operators the naming is worst for,
because a name needs a block whose first reporter the feed could separate and
they win few blocks.
The preimage is derived from a secret only its owner holds, so the same
preimage on two chains is the same operator. That is an inference the site can
stand behind, unlike guessing from timing — and it is the only lever that works
without waiting on the feed at all.
It is a separate `AttributionSource` and renders as one. `Carried` names carry
no confidence and no votes, because they assert nothing about this chain: only
that this is who the operator was last known to be. Any live attribution here
outranks one, and the row shows an `elsewhere` chip rather than the telemetry
`node` chip so the two can never be read as the same claim.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
A mapping outlives the votes that earned it — deliberately, so a name does not
blink out whenever the feed goes quiet. But the tooltip then rendered "0% of 0
attributed blocks point here", which is arithmetic nonsense dressed as a
statistic.
The top miner takes 87% of blocks at a two-second interval, so its twenty
observations span about forty-seven seconds; votes age out of the window
constantly and this is the ordinary state, not an edge case. It now says the
name rests on evidence older than the current window, which is what is true.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
A name now appears as soon as one node leads the vote, rather than waiting for
three. On a chain that resolves a first reporter for about one block in seven,
the old bar left almost every miner permanently anonymous — and since a later
block can overturn a mapping, a guess that corrects itself is more useful than
a blank that never fills in.
What is refused is a dead heat. Choosing between two equally-voted nodes would
be a coin toss wearing a telemetry badge, which is a different thing from a
thin but real lead.
The honesty moves to the presentation instead of being dropped. A percentage
alone cannot separate "100% of 1" from "86% of 7", so `Attribution` and
`LeaderboardRow` now carry the number of votes behind the name. The leaderboard
marks a mapping resting on a single block or a narrow lead with `node?` and a
dashed chip, and the tooltip states both figures and that a later block can
correct it. Dashed rather than a second colour: the accent hue is reserved for
identity and alarm, and this is neither — the same claim held more loosely.
The footer now says node names are inferred from which node reported a block to
telemetry first, never something the miner asserted, and that the reward
preimage beside them is the only identity the chain itself vouches for.
`a_single_lucky_vote_still_names_nobody` asserted the old policy and is
replaced by `one_vote_names_but_says_it_is_only_one`, which pins the new
contract: it names, and it reports `votes == 1` so the UI can mark it. A new
`a_dead_heat_names_nobody` keeps the case that is still refused.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Reported from the live site: a name would appear and then be "often forgotten
if it doesn't mine a new block within the window".
`attribute` dropped the held key the moment the vote window contained no vote
for it. That reads silence as contradiction — the same mistake as counting
abstentions in the confidence denominator, one layer further on. The window
holds twenty observations and mainnet resolves a first reporter on roughly one
block in seven, so an incumbent's votes routinely roll out before any
replacement earns one. The name vanished at that moment and the row fell back
to the raw preimage until three fresh votes rebuilt it.
A held name is now replaced rather than dropped: it stands until a different
node leads it by `TAKEOVER_MARGIN`. The margin is the other half of the same
problem — a window carrying about three cast votes flips on a single vote under
a bare majority rule, and pools demonstrably run several nodes that all report
their blocks (this miner has answered to both `QUANPOOL - quanpool-com` and
`quanpool-payout-mainnet`). Requiring a lead of two rides that out instead of
oscillating between siblings.
Takeover after a restart or rename still works, and is still tested: a node
that genuinely owns the blocks accumulates votes and passes the margin within a
few blocks.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Names were still absent after fixing the confidence denominator, because no
vote was being cast at all. The cause is that `ATTRIBUTION_SETTLE` was doing
duty for two unrelated quantities.
It was written for one: how long node reports take to spread across the network
once telemetry has the block. Eight seconds is right for that. But the lookup
also has to wait out a second quantity nobody had measured — how far behind the
chain the feed itself runs. On Planck the two coincide, because ~240 nodes at a
13 s block time is ~18 reports a second and the feed keeps up. Mainnet is ~190
nodes at ~1 s, an order of magnitude more traffic, and the feed sits about 57
blocks back. Measured against the node: chain tip 9539, highest height the feed
had reported in the preceding 110 s was 9482.
So every lookup asked who reported a block roughly fifty seconds before
telemetry had heard of it. `first_import` missed every time, every observation
became an abstention, and no author was ever named.
The join now retries every 5 s until the feed reaches the block, giving up at
three minutes and recording an honest abstention then. It adapts to whatever
the lag is rather than assuming it is zero.
Confirmed live: the top miner resolved to `QUANPOOL - quanpool-com` at
confidence 1.00 within about two minutes of restart, from 3 votes cast over 21
blocks. That 3-of-21 is also why the previous commit was necessary and not
sufficient — under the old denominator it would have scored 0.15 against a 0.6
bar and stayed nameless.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Mainnet has shown no node names since launch. Every miner renders as its raw
preimage while Planck, on the same code and the same feed, names twenty.
`votes` holds one slot per block checked and a `None` for every block whose
first reporter the feed could not separate from its second. Those are
abstentions — the feed did not say who led — but confidence divided agreement
by `votes.len()`, counting them as votes against. Measured on the live feed,
mainnet resolves a clean lead on about a quarter of blocks, so the best score
any author could reach was 0.25 against a `MIN_CONFIDENCE` of 0.6. No name
could ever be held, however unanimous the votes that were actually cast. Planck
cleared the bar only because 41-87% of its blocks produced usable votes.
Confidence is now agreement among cast votes, which is what it always meant:
"of the blocks where we could tell who was first, this fraction pointed here".
`MIN_VOTES` guards the other side, so a single vote in an otherwise silent
window is not read as 100% agreement.
`a_held_name_survives_a_confidence_dip` asserted `confidence < 1.0` after five
unresolvable blocks. That assertion encoded the bug. It now asserts the
opposite for abstentions and keeps a real dip — disagreeing votes — as the case
that lowers confidence.
Two false leads are recorded in CLAUDE.md so the next person does not spend the
afternoon on them: the feed delivers a 190-node roster before any live block
traffic, so `ImportedBlock` does not appear for ~40 s after subscribing and a
short probe concludes the feed is dead; and the roster's `AddedNode` entries
carry a height-0 placeholder block with a zero hash, which inflates the
apparent count of nodes claiming to be first if counted as import reports.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Mainnet launched today and the node was switched to it. The observer reads
whatever the node is on, so the `[[chains]]` entry pointing at 127.0.0.1:9944
had to follow it — until now the daemon was filing mainnet's blocks under the id
`planck`, with Planck's genesis and PLK still attached to them.
Read from the node, not assumed: `system_chain` "Quantus", genesis
`0xfb5487c0…626fba`, token QTC. `mainnet = true` is asserted here rather than
inferred, which is the field's whole purpose — telemetry also carries a chain
calling itself "Quantus Staging Mainnet", and it is not this one.
The target block time is **12 s, not the 6 s Planck used**.
`TARGET_BLOCK_TIME_MS` is 12_000 in the runtime the node actually runs —
checked at commit b017e642, the `0.11.1-b017e6420aa` its `system_version`
reports, rather than at the tip of the chain repo. It is the denominator for
every hashrate published before twenty tip samples exist, which on a chain
hours old is all of them, so carrying 6.0 across would have put out a mainnet
hashrate at exactly twice its true value — labelled "nominal", not broken.
Planck keeps its 268 telemetry nodes and stays in the site's navigation as
`no_endpoint`: still listed, no longer navigable, because authorship comes from
block headers and we no longer run a node that serves Planck's. That is an
ordinary state here, not an outage.
Nothing about the node itself is touched by this; every command run against it
was a read.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
`#/:chain/:height` opens a block and then rewrites the address to
`#/:chain/:hash`. The redirect is the point, not tidiness: `block` is keyed on
(chain, height) and a reorg overwrites, so a height names a *position* and a
link to one comes to mean a different block the moment the chain forks there. A
hash names one block for good — including after it has lost, which is the only
way an orphan is linkable at all.
`GET /v1/chains/{chain}/blocks/{ref}` takes either spelling; a decimal number
and 32 bytes of hex cannot be confused, so a caller holding a number is not made
to guess which one this API wanted. The same reasoning puts block refs in the
second path segment beside the window: no window name is all digits or 0x plus
64 hex characters.
The answer is assembled from two sources because neither is sufficient, and
they fail at different times:
- The node holds the block and forgets it. Non-canonical bodies go once finality
passes them (blocks-pruning defaults to archive-canonical, and Planck finalises
~100 blocks back), and difficulty is a state read behind a 256-block default.
- The observer holds what the node never had — when the block was seen, whether
that sighting was at the tip — and what it has since forgotten: the difficulty
read while the state behind it still existed.
So resolution degrades in a stated order rather than failing. Node first; a
header it cannot serve falls back to `block_displacement`, which is what makes
an orphan describable at all; a node that is away falls back to the recorded
row. `from_node` tells the reader which they are looking at, and every value
neither source has is spelled out in words — "the body is pruned", "this
observer never saw this block" — because a blank on a block page reads as zero.
Verified against the live chain: a tip block carries difficulty, one extrinsic,
a 1.9 s gap and 473 ms of propagation; height 1000000 keeps its body and loses
difficulty to state pruning; genesis has no author digest and does not panic; a
seeded displacement serves as an orphan with the winner linked, and the winner
lists it in `also_seen`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
`block` is keyed on (chain, height) so a replacement overwrites rather than
accumulates — a schema keyed on the hash would inflate the losing fork's author
forever. The cost is that the losing block vanishes without a trace, and with it
any record that the height was ever contested. `block_displacement` is that
trace: one row per displacement, holding the whole row that was about to be
overwritten.
It cannot be reconstructed later. `blocks-pruning` defaults to
archive-canonical, which discards non-canonical bodies once finality passes them
— about a hundred blocks behind the tip on Planck — and difficulty is a state
read against a 256-block default. Minutes after a reorg the chain itself can no
longer answer for the block, so what is captured at the moment of the swap is
all there will ever be. That is why the row holds the author and the difficulty
rather than just the hash: the author is the interesting field, it says who lost
the race, and it costs a table that gains a row only when the chain forks.
The capture is a CTE ahead of the upsert, in the same statement. Every
data-modifying CTE sees the same snapshot, so it reads the pre-update row even
though the insert below it is replacing that row in the same breath, and a
failure rolls back both — which two statements without a transaction would not.
A trigger would do the same thing invisibly to anyone reading the query.
Nothing reads the table yet, deliberately. The read shape belongs with the
block route that will use it; what matters now is that the history exists by
the time that lands, which it cannot do retroactively.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
`cargo fmt --check` has gated the Rust half since the first commit; nothing
gated the TypeScript, so 25 of 31 files under web/src had drifted from the
.prettierrc that was already sitting in the repo. `pnpm format:check` is now the
counterpart, in the same step as `pnpm lint`.
The reformat is whitespace only — the emitted bundle hashes identically before
and after.
web/.prettierignore excludes src/api/generated. Those files are written by ts-rs
from the entities crate and CI already fails on any diff in them after `cargo
test`; formatting them here would be reverted by the next regeneration and turn
that drift gate red on every subsequent build. Verified by regenerating and
confirming the tree stays clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Height is the one tile with no sparkline — it only goes up, and a straight
diagonal reports nothing — so it sat in the middle of the row with an empty
slot under it, splitting the four lines into two and two. Last, it stops
interrupting them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Adds `GET /v1/chains/{chain}/series` and a sparkline under network hashrate,
difficulty, block time and miners. Height gets none — it only goes up, and a
straight diagonal reports nothing.
The chart is bucketed by height, not by time: every point is the same number of
blocks, so every point carries equal statistical weight and a stretch nobody
recorded stays the same width on the axis as one that was. A bucket is the
window over fifty, which makes a full window exactly fifty buckets — that is
what lets the rolling miner count at the last point be the same number printed
in the tile above it, rather than a per-bucket count that would sit well below
it. Buckets are anchored to absolute height for the same reason `miner_series`
bins from the epoch: anchored to the tip instead, every boundary slides by one
on every block and the chart shifts under a reader watching it.
Interval comes from `authored_at`, never `observed_at`. A gap fill writes a
whole batch of observation times within the same second, so deltas taken from
it read as a chain producing hundreds of blocks a second — the trap
`RollingWindow`'s `at_tip` flag exists to avoid, which until now lived only in
memory. It is a column as of `0002`, so the distinction survives a restart and
`observed_at` is interpretable at all.
Each point's interval is a three-bucket moving average. The mean of one
bucket's gaps still carries real Poisson error — 72 blocks of a 15 s interval
lands at ±1.8 s — and across fifty points that draws a chain that appears to
change size every few minutes and does not.
Fixes a wrong number found on the way in: `pallet_qpow` retargets in
`on_finalize`, so `QPoWApi_get_difficulty` at the latest state is the
difficulty for the *next* block, not the one just seen. Every stored difficulty
was off by one block — invisible, because a gradual retarget still looks
plausible. It is now asked at `header.parent_hash`, the same call the miner
made when it built the block. A node that has pruned that state gets `None`
rather than the current value: substituting it is exactly how a gap fill would
stamp today's difficulty across a stretch of old heights. The CLI backfill had
the same bug at greater scale, copying one reading across an entire range.
Unknowns are drawn as breaks in the line, never interpolated: an outage, a
difficulty a pruned node cannot supply, a point without a full window of
authors behind it. A chart that guesses across what was never recorded is the
same class of lie as a hashrate divided by a sync speed, and just as hard to
catch afterwards.
Read against the dataviz skill first, per CLAUDE.md. Single series, one hue;
emphasis on the newest point is `--data-bright`, a lighter step of the same
bronze, because the usual "current period in the accent" convention is the one
change that would break this palette. The hover readout borrows each tile's
note line — five floating tooltips across a row of 180px tiles is a pile, not a
hover layer — and the same readout is on keyboard focus and in the aria-label,
since the line does not start at zero and its height is not a quantity anyone
should be estimating.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jp6a8EDar9ueEhAxzep4V5
Each of these failed a CI run this session and none of them looks like what it
is: a loopback health probe refused by a healthy daemon, a proxy that cannot
resolve its own public name, and 'cargo: command not found' on a node runner.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
The probe curled 127.0.0.1:25864 and got connection-refused against a daemon
that was running perfectly — the journal it captured on failure showed it
serving. The API binds the host's *mesh* address by design: the fleet has one
firewalld default zone, so a wildcard bind plus the named service would publish
it on every address the host carries, and the edge proxy is at another site, so
loopback-only would not serve the site at all.
`hostname -f` resolves to that mesh address, so the probe needs no value carried
between steps. Verified against the running service: loopback refuses, the FQDN
returns 200.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
Three things that would each have failed the run seconds in
(architecture/gitea-runners.md §4):
- `runs-on: fedora-43` for a cargo build is `cargo: command not found` — the
fedora images are node/shell runners with no Rust. `rust` is built on
`runner-fedora-44`, so it carries node + pnpm as well and one job still builds
both halves. musl stays: that runner is Fedora 44 and the targets are 43.
- `corepack enable` is not bundled by any image and pnpm is already on PATH.
- The site check curled the public name *from the edge proxy*, where it resolves
to the site WAN address and dead-ends on the OPNsense LAN interface. Pinned to
loopback, which still exercises the stream router, SNI, the vhost, the
cross-site hop and the API.
Deploy jobs stay on `infra`: §4 is about tooling, not routing, and these targets
are mesh-only .internal names — same reasoning as lair/quantus and lair/mail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
`CorsLayer::allow_origin` replaces rather than appends, so folding over the
configured list left only `blackbeard.internal` allowed and silently refused
`blackbeard.observer` — the site's own public origin. It broke nothing, because
the frontend is served same-origin and never consults CORS, which is precisely
why it would have gone unnoticed until something else called the API.
Verified on the deployed vhost: both configured origins are now echoed back and
an unlisted one is refused.
Also records the deployment gotchas this session turned up (exact-argument
sudoers matching, the runas spec for the config check, the cross-site hop the
loopback probe cannot see, and why a WebSocket upgrade test needs --http1.1).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
The closing heredoc interpolates $EDGE_HOST, so backticks in its prose ran as
command substitution ('create: command not found').
A missing psql on the app host was reported as an authentication failure, which
sends you looking at the CN mapping instead of at a package that has no reason
to be installed. An unreachable standby is now a loud, specific warning: pg_ident
is not replicated, so a failover to a standby that never got the mapping locks
the app out during an outage.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
The chain list is no longer config. substrate-telemetry announces every chain it
knows — with live node counts — before and independent of any subscription, so
one feed connection that never subscribes is a complete, self-maintaining index.
A chain that launches tomorrow appears in the nav on its own.
Config now contributes only RPC endpoints. Authorship lives in the `pow_` digest
of a block *header*, and telemetry publishes hashes and heights but never
headers — so a chain with no endpoint is listed with its node count and is not
navigable, and the nav says why rather than offering an empty board. The two
sets join on genesis hash, never on name.
Two chains are tracked in full: Planck (our own node on bob) and Heisenberg
(a1-heisenberg.quantus.cat, found via Quantus-Network/quantus-apps). Both decode
identically — same digest shape, same QPoW runtime API — which is the proof that
adding a chain is one [[chains]] entry and no new code. Quantus Staging Mainnet
and Quantus Dirac Testnet are listed and disabled: no endpoint exists under the
a1-/a2- pattern and none resolves under any plausible name.
Nav is ordered by node count descending, ties broken on id so it cannot reshuffle
between refreshes.
Telemetry now also yields chain-wide best height, best finalized height, average
block time and the client-version histogram (feed codes 1, 2, 12 and 22).
Infra moves to a single site: oolon fronts blackbeard.observer, reverse-proxying
cross-site over the mesh to the API beside the node on bob. That hop is the one
the loopback health probe cannot see, so the deploy now checks it explicitly —
a firewalld service scoped to bob's own /16 would leave a live site with a dead
/v1 and nothing would fail. infra-setup gains `dns` and `cert` roles that run on
oolon with the credential already there: the Cloudflare token is never copied off
the proxy, and it refuses to repoint an apex record that already exists.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
Cargo workspace plus a Vite frontend, following ~/git/architecture/generic.md.
Every block header carries its author's wormhole reward preimage in a `pow_`
PreRuntime digest, so authorship for the whole network is derivable from headers
alone — no indexer, no registration, no way for a miner to be left out. That
decoding, the hashrate maths and the telemetry name attribution live in
blackbeard-core with no I/O at all, so the parts that are easy to get subtly
wrong are exercised by unit tests rather than only against a live chain.
The browser holds one WebSocket: snapshot on subscribe, deltas thereafter. The
head stream is itself a push (chain_subscribeNewHeads), so a block reaches the
page the moment the node imports it. Messages are serialised once per broadcast,
and leaderboards are recomputed only for windows a socket is actually watching.
No RxJS — useSyncExternalStore is React's own contract for this.
Verified against the live Planck testnet: 12/12 headers decoded, telemetry names
attributed (quanpool-planck, baba-gorchitsa, …), warm start restoring 84 blocks
and 5 held names across a restart.
Three findings worth recording, all in CLAUDE.md:
- substrate-telemetry sends its JSON in *binary* frames. A text-only client
connects, subscribes, reports healthy and receives nothing at all — and a
Python probe hides it, because json.loads accepts bytes.
- Difficulty is a little-endian U512; decoding it big-endian gives a number
wrong by ~10^150 that still renders fine.
- Planck's real block interval is ~13-15s against a 6s target with enormous
variance, so a measured interval needs 20 tip samples before it is publishable.
Deploy assets, the Gitea Actions workflow and script/infra-setup.sh are included;
port 25864 is registered in architecture/port-allocations.md.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MSDYiibCtELsrjQq6KXnoi
