Sighte Addons

The clear is a duration now, and rooms are weighted instead of counted.

What changed

• #31 residue-001 — a floating-point residue (3.55e-15 unattributed points) no longer reaches permanent telemetry. It was clamped on one side only; both sides are settled now.
• #32 clear-001 — the report schema goes 4 → 5. enterTick is stamped when a party member has actually stayed in a room, not the first time anyone is glimpsed in it. A walk-through no longer reports as that room's clear time. On a real M7 floor the anchor landed within one second of entry in nine rooms out of ten; the tenth cleared in six ticks and used the bounded fallback.
• #35 / #37 clearpoints-001, clearpoints-002 — points are weighted rather than counted. A room is worth 0.75 plus 0.25 per secret, with puzzles seeded higher (Ice Fill 2.0, Water Board 1.5), and that base moves toward the room's measured clear time as observations accumulate. Kind and size are no longer hand-set constants — a puzzle scores high because puzzles measure slow.
• #39 chat-001 — deaths are charged on the tick chat announces them instead of up to a second later via the tab list. Wither doors, blood doors, puzzle solves and puzzle failures are recorded for the first time. Hypixel names a finder only for wither essences, so that is the only secret chat can attribute; everything else keeps the existing inference.
• #41 runloss-001 — a run quit from inside a dungeon is no longer lost. The report was written only when a new world was joined, so closing the client from a floor discarded it. This cost a real run during testing: ten cleared rooms and 24 secrets, gone.
• #40 party-001 — no behaviour change. The decoration-to-player assignment was made testable, and the upgrade path recorded in the source for a long time was shown not to exist: the map decoration keys do not survive the wire, and the client invents them from list order.
• #38 artifacts-001, #33 — a real dungeon run committed as evidence, and init.sh no longer tells every session to run a command that overwrites the released jar.

What is not in this jar

The receiver in Sighte/skyblock-server is already deployed and ahead of this build — nothing here is waiting on the box:

• schema-001 (PR #13) routes v5 clears into their own average, so the new anchor never mixes with the old one.
• scores-001 (PR #19) recomputes room scores every half hour on a systemd timer.
• scores-002 (PR #20) serves them at GET /roomstats.

For the operator: nothing to do. All three are live on master and verified on the box.

What this changes for older installs

Nothing breaks. The receiver validates v as 1..10 and has accepted enterTick since schema 3, so a 0.9.0 client keeps uploading successfully and its clears keep folding into the pre-v5 average. There is no rejection and no data loss for anyone who does not update.

What a 0.9.0 install misses is the corrected clear anchor — its clear times stay upper bounds.

Requirements

sha256 f4721e9ef368abfa869af147caabd3e8baa6334aa4a779aac2dfc6ee3adc9bd2

What is not verified

The development client cannot reach Hypixel, so several paths in this build are reasoned from source and bytecode rather than observed in a real dungeon:

• runloss-001's trigger. That Fabric raises DISCONNECT on a real quit is read out of the call graph — Minecraft.destroy → Connection.disconnect → Netty's channelInactive — and has never been seen. The write path itself is tested.
• chat-001's line patterns are taken from the sources of published mods (SkyHanni, Cowlection, IllegalMap, OdinLegacy) and checked character-for-character against them, but Hypixel has not been observed emitting them. A wrong pattern matches nothing and leaves the existing behaviour unchanged.
• Whether a chat secret line arrives before the action-bar update it attributes. If it arrives after, that half does not work — and widening the window is not the fix.
• The measured half of the scoring model. Every published score today is exactly its seed, because no schema 5 build had shipped until this one. Runs from 0.10.0 are the first that can fill it.
• Party attribution. The decoration-order heuristic has never been exercised against a real party; the one real run used for testing was solo.