Potential Bugs Backlog¶
A queue of incidental discoveries — suspected bugs, latent races, missing validations, stale comments — that were spotted while working on an unrelated task and deliberately not fixed in-line, per the Stay-On-Task Policy in .junie/AGENTS.md.
This file is a backlog, not a tracker. Promote an entry to a Linear issue (RTP-<n>, see Issue Tracking (Linear) in .junie/AGENTS.md) or fold it into a future task's Effective Issue when it is ready to be worked on. Once filed, the entry keeps its technical detail and gains a **Linear:** line; Linear owns its status from then on. Once an entry is resolved, delete it — this file does not maintain a resolved-bug archive.
What this file is — and is not¶
Yes: "I was doing X, I noticed Y looks broken, Y is not part of X, and I am walking away from Y. Recording it here so a future task can pick it up."
No — do not use this file for any of:
- Work you are doing or just finished as part of the current task. Use the
<UPDATE>checklist, thesubmitsummary, the commit message, andCHANGELOG.mdfor user-visible changes. - A diary of your own fix attempts, build outputs, packaging chains, or per-session follow-ups. If you opened the entry and resolved it in the same session, delete the entry — it never belonged here. Do not annotate it with
**Resolved:**/**Follow-up:**bullets. - Durable engineering lore or repro recipes →
LESSONS_LEARNED.md. - Roadmap items or deferred design → the relevant plan doc or an ADR.
- Session resumption state → your
<UPDATE>checklist ordocs/dev/scratch/CHECKLIST-<slug>.md. - Test failures or CI noise from the current change → fix them or escalate; not here.
A correct entry describes someone else's future problem that the current task is choosing not to solve. If you catch yourself writing a multi-paragraph resolution log on an entry you authored this session, that is the misuse signature — remove the entry instead.
How to add an entry¶
Append to the Open section below using the template. Keep entries short — one paragraph each. If a deeper analysis is warranted, link to a separate doc rather than inlining it here.
Entries in the Open section are ordered by priority (highest first): runtime crashes and safety/thread-safety hazards first, then correctness/maintainability and operator-facing config issues, then performance, and finally cosmetic / log-noise / test-noise findings. When adding a new entry, insert it at the position matching its severity rather than strictly by date.
Template¶
### YYYY-MM-DD — <short title>
- **Severity:** Critical | High | Medium | Low | Cosmetic
- **Status:** Open | Triaged | Under Investigation | In Progress | Blocked
- **Discovered during:** <issue ref / short task description>
- **Location:** `<path/to/File.java>` line <N> (or symbol name)
- **Symptom / hypothesis:** <one or two sentences>
- **Impact:** <user-visible effect, best guess>
- **Suggested next step:** <minimal investigation or fix sketch>
- **Linear:** RTP-<n> (add only once filed; omit before)
Open¶
2026-11-08 — lcv_vlc_demux_open calls lcv_sys_index_file_superblocks with stale 10-argument signature¶
- Severity: High
- Status: Open
- Discovered during:
:leaf-videoADR/LICENSE/build portability cleanup (:leaf-video:compileAndRunNativeGate1) - Location:
resource/LeafVideo/native/lcv_codec_system_plugins.clcv_vlc_demux_open(line 1781) vs.lcv_sys_index_file_superblocks(line 235) - Symptom / hypothesis:
lcv_sys_index_file_superblockswas refactored to take 7 arguments(FILE *fp, uint64_t file_size, uint64_t *out_offsets, size_t *out_lengths, uint8_t *out_k_vals, uint8_t *out_first_hdr64, size_t max_sbs)(as used bylcv_probe_stream_fileandlcv_decode_file_frame_rgba8), butlcv_vlc_demux_openstill passes 10 arguments(psz_file, &sb_offs, &sb_lens, &sb_ks, &sb_count, &w, &h, &fps, &fps_num, &fps_den). - Impact: Native compilation of
resource/LeafVideo/native/lcv_codec_system_plugins.cfails in:leaf-video:compileAndRunNativeGate1and:resource:LeafVideo:compileNativeAccel. - Suggested next step: In
lcv_vlc_demux_open, openpsz_filewithfopen(psz_file, "rb"), determinefile_sizeviafseeko/ftello, allocatesb_offs/sb_lens/sb_ksup toLCV_MAX_SCAN_SUPERBLOCKS, call the 7-argumentlcv_sys_index_file_superblocks(fp, (uint64_t)fsz, sb_offs, sb_lens, sb_ks, hdr0, LCV_MAX_SCAN_SUPERBLOCKS), and parsew,h,fps_num,fps_denfromhdr0+lcv_sys_resolve_tracks_or_obs_fps.
2026-11-14 — Untracked LcaStreamingNbtEncoder calls missing LcaLiveArchive package-private accessors¶
- Severity: High
- Status: Open
- Discovered during: LeafGen tree/biome/path/cave worldgen verification (
:resource:LeafGen:leafgen-common:test) - Location:
api/anvil-api/src/main/java/io/github/dailystruggle/anvil/LcaStreamingNbtEncoder.javalines 91–93, 114–116, 127, 320–322, 357 vs.LcaRegionWriter.LcaLiveArchive - Symptom / hypothesis:
LcaStreamingNbtEncodercallsarchive.minSectionY(),archive.sectionCount(),archive.defaultDataVersion(),archive.isWriteModeBZstd(), andarchive.encodeModeBWithAtomicDictForStreaming(chunkFrame, headerLen), which are not yet exposed onLcaRegionWriter.LcaLiveArchivein the local working tree. - Impact:
:anvil-api:compileJavafails with 10cannot find symbolerrors unless those package-private helpers are wired onLcaLiveArchiveor:anvil-api:compileJavais skipped against prebuilt:anvil-apiclasses. - Suggested next step: Expose package-private
minSectionY(),sectionCount(),defaultDataVersion(),isWriteModeBZstd(), andencodeModeBWithAtomicDictForStreaming(byte[], int)onLcaRegionWriter.LcaLiveArchivein theleafanvil-ADR-005workstream and run./gradlew :anvil-api:test.
2026-10-08 — Probe governor's probe time includes coalescer queue wait¶
- Severity: Medium
- Status: In Progress
- Discovered during: stress run
20261008-170548([RTP][probe-gov]lines:Tp = 10.2 msvsTf|direct = 8.5 msatp = 0.073) - Location:
rtp-core/.../selection/region/ProbeFirstGovernor.java(sample recording) and the probe call sites inPregenTask/QueueTask - Symptom / hypothesis:
Tpis measured from dispatch to result, so a probe parked in the per-region-file coalescer counts idle wait as cost. In skip mode the 1-in-32 trial probes rarely join a group and each pays a full open, so the losing option is sampled at its worst (ADR-109 measurement section). - Impact: The fill path can settle on skipping the probe for the wrong reason on fresh worlds, where same-file grouping is worth the most.
- Suggested next step: Fix pending bench verification: probes carry their drain's per-chunk share and
groupSize(ColumnProbe/ChunkColumnProbe.groupSize());ProbeFirstGovernorfitsTp(g) = perChunk + open/gand evaluates it at the default-probing group size; SKIP-mode trials run up to 8 candidates of one region file (binKey);ScanTaskcompletion resets the region's governors (reload already did). Covered byProbeFirstGovernorTest.groupedTrialsJudgedAtDefaultGroupSize. Remove this entry once a fresh-world stress run shows[RTP][probe-gov]Tpnear the PROBE-mode value while in SKIP (and a flip back wherepwarrants it).
2026-10-08 — StressTestRTP spark thread selector matches no thread; saved profiles are empty¶
- Severity: Medium
- Status: In Progress
- Discovered during: Linux A/B stress runs
20261009-005522and20261009-010359(spark async mode, "save complete" logged) - Location:
helpers/StressTestRTP/.../SparkHook.javabuildStartCommand(~175-178); defaultspark.threads: "Server thread,RTP-Anvil-IO"inconfig.yml - Symptom / hypothesis: The comma list is appended as a single
--thread Server thread,RTP-Anvil-IO. Spark re-joins space-split tokens (the space is harmless) but takes the whole comma list as one name, so no thread matches. The.sparkprofileholds only server metadata (config, plugins, datapacks) and no thread samples. The pool threads are namedRTP-Anvil-IO-<n>, so an exact name would not match them either. - Impact: Every profiled run since the selector default was added has produced no samples;
thread_cpu_topin the summary is empty. Runs remain valid as unprofiled benchmarks. - Suggested next step: Fix pending bench verification:
buildStartCommandemits one--threadper entry (unquoted; quotes would become part of the name) and adds--regexwhen an entry has a*glob, since spark resolves exact names to thread IDs once at start and would miss pool threads spawned later. Default is nowServer thread,RTP-Anvil-IO-*; covered bySparkHookTest. Remove this entry once a saved profile shows non-zero thread nodes for both threads.
2026-10-06 — ACCUMULATE repeats used chunks; landing spacing loosens on dense learned bad area¶
- Severity: Medium
- Status: Triaged
- Discovered during: cross-plugin destination chart for the Paper 26.2 stress run
20261006-135938 - Location:
MemoryShape.resolve(ACCUMULATE good-index remap; uniquePlacements gated onexpand);AbstractDualLayerShape.selectL3Candidate/keyForCounter(BINNED_AMORTIZED) - Symptom / hypothesis: In teleport order, 79 of LeafRTP's 4,096 landings fell within 256 blocks of one of the previous 64 (35 by the same player), against about 69 for a random scatter. Modelled by
ConsecutiveLandingSpacingSimTeston the shipped default circle (P=32, S=1024 now that the stride cap is the full bin, ADR-088). The raw draw order keeps spacing: no lane holds two keys under 256 blocks apart, and there are 0 back-to-back pairs under 256 blocks per 4,096 draws (uniform 0.88). Per 4,096 landings with ~40% synthetic bad terrain (uniform: back-to-back 1.13, 64-lookback 64.1, repeats 3.94): - Spacing holds under rejection and cache mixing: cold-only ACCUMULATE gives back-to-back 0.06 and 64-lookback 50.0; FAST/COLD interleaving gives 0.06 and 54.4. The remap offset between two nearby picks is only the rejections learned between them, a few Hilbert steps, so the P-chunk lattice is nudged, not lost; the exact stride residue (0.5% kept) is stricter than spacing. FAST and COLD share one
rand()stream, so mixing adds no repeats (3.56 against 4.44 cold-only). The 64-lookback metric is not a lane promise: windows of 16-64 draws change lane. - Repeats under ACCUMULATE (known limit): every merged rejection renumbers good indices, so a used chunk can reappear at an undrawn index; with
expand: falselandings are never marked, so repeats sit at the uniform rate (4.44 vs 3.94, about 0.1% of landings). REROLL never renumbers. Marking landings is gated onexpandon purpose: without expand the range never grows back, so uniquePlacements marks drain the good domain until search tasks pile up without finding a location (MemoryShape.resolve). - Dense learned bad area (open, not simulated): where scan or pregen has learned large bad regions, the offset difference between nearby picks approaches a full bin (1,024) and in-bin spacing falls back toward uniform.
- BINNED_AMORTIZED (open): this non-default backlog runs its own counter, stride rule (capped at 256, no phase windows) and per-draw phase, so it collides with the main ordering and comes out worse than uniform (1.19x).
- No model reproduces the stress run's excess over uniform (79 vs ~69); real terrain clustering is the likely cause.
- Impact: With
expand: false, consecutive players can land on a chunk an earlier player used, about as often as with random picks. Back-to-back spacing holds on typical terrain and loosens where learned bad area is dense. Disclosed indocs/admin/configuration/REGIONS.md(Spacing, repeats and worst case by mode); FRONT_PAGE.md links there. - Suggested next step: Keep the repeat limit documented; both obvious fixes are ruled out. Physical-order lanes need per-lane occupancy state instead of run tables (the encoding that failed earlier performance tests, and it breaks C8), and marking landings without
expandexhausts the region. Any future fix must consume no good area and keep the run encoding. Route BINNED_AMORTIZED through the main counter and phase windows. Flip thedups > 0pin inConsecutiveLandingSpacingSimTest.terrainRejectionKeepsBackToBackSpacingonly with such a fix. - Linear: RTP-2
2026-10-06 — Weekly Jazzer job fuzzes one Anvil target for 2 s and passes¶
- Severity: Medium
- Status: In Progress
- Discovered during: v3.3.0 pre-release audit (Jazzer fuzz targets)
- Location:
.github/workflows/fuzzing.yml;api/anvil-api/src/test/.../AnvilRegionFuzzTest.javaand itsAnvilRegionFuzzTestInputs/seeds; rootbuild.gradletasks.withType(Test) - Symptom / hypothesis: Jazzer 0.30.0 fuzzes one
@FuzzTestper run, the old workflow set a property the forked test JVM never saw, and a cached:anvil-api:testresult could skip fuzzing entirely. - Impact: The fuzzing badge gave false assurance until the workflow ran each target for its configured duration.
- Suggested next step: Fix pending the first scheduled CI run: the workflow runs one job per
@FuzzTestmethod withJAZZER_MAX_DURATION, the rootTestconfig declares the Jazzer env vars as inputs and disables up-to-date checks in fuzz mode, andfuzzReadChunkView/fuzzReadColumnProbecarry real-region seeds (local 15 s runs of both passed). Remaining: seeds for the LZ4 and NBT targets (needs extracted payloads); crash reproducers go under<Class>Inputs/. Remove this entry once a CI run shows all five jobs fuzzing for the set duration. - Linear: RTP-21
2026-10-18 - Web editor paints whole walk-path squares red when hazard runs exceed the cap¶
- Severity: Medium
- Status: Open
- Discovered during: user report of solid red squares on the walk path (SQUARE region, radius 830c, centerRadius 121, P=16, Paper 26.2)
- Location:
rtp-core/.../commands/editor/EditorCurveModel.javaMAX_HAZARD_RUN_BYTES;EditorSessionManager.generateVisualizationPayload(hazardRunsSkipped: "hazard-cap");EditorLiveFeed/HazardTiles;docs/editor/index.htmlpathHazardSampler - Symptom / hypothesis: 100,140 bad runs exceed the 256 KiB cap, so the page colours the path from the export-time 2D grid (128x128 cells, about 13 chunks each); one bad chunk turns a whole cell red, and live hazard updates stop. Curve, scan data and
ScanTaskchecked and ruled out; at 512c the exact runs fit and the render is correct. - Impact: Misleading solid-red blocks in the editor on large regions; display only, selection unaffected.
- Suggested next step: Ship hazard runs coarse to fine like biomes (
BiomeBinCodeclevels 32/4/1, finer levels bin by bin via a level field onhazard-delta, bad share per stretch). Stopgap: raise the cap to about 1 MiB. - Linear: RTP-72
2026-10-14 - Folia arrival platform writes blocks in one unbudgeted destination-region task¶
- Severity: Medium
- Status: Open
- Discovered during: dividing Folia region-thread work per region (ADR-112)
- Location:
platforms/folia/folia-common/.../entity/FoliaRTPPlayer.javateleportAsync(...).thenCompose(~188-228);FoliaRTPWorld#platform(~1029-1105) - Symptom / hypothesis: After
teleportAsynccompletes on the destination region, the emergency platform is built inline with nestedrunTaskper chunk andsetTypeloops sized byplatformRadius. It is player-facing, so the new per-region budget only charges it. - Impact: Many simultaneous arrivals in one region with large platforms can still extend that region's tick.
- Suggested next step: Measure with the stress harness; if it shows, split block writes across ticks per chunk or cap writes per tick.
- Linear: RTP-93
2026-10-14 - LocationGenerator.joinSafely waits on work queued to the same pool¶
- Severity: Low
- Status: Open
- Discovered during: bounding the Fabric/NeoForge async pools
- Location:
rtp-core/.../selection/region/LocationGenerator.javajoinSafely - Symptom / hypothesis: Blocks a pool thread on a future whose work is submitted to the same bounded pool. Only tests call it today, and it has a timeout.
- Impact: None in production; a future production caller could stall a bounded pool until the timeout.
- Suggested next step: Remove it or make callers compose the future instead of joining.
2026-11-04 - LeafFuzz: Detect rebranded/stolen JARs and ZIP-path vs bytecode this_class mismatches in batch triage¶
- Severity: Low
- Status: Open
- Discovered during: empirical 453-JAR Modrinth batch triage validation (
Dynamics GUI-1.0.jar,ItemTeleporter (1).jar) - Location:
api/bytecode-api/src/main/java/io/github/dailystruggle/bytecode/JarConfigScanner.java;api/bytecode-api/src/main/java/io/github/dailystruggle/bytecode/BytecodeCausalGraphScanner.java;resource/LeafFuzz/leaffuzz-common/src/main/java/net/leaf/fuzz/common/cli/LeafFuzzBatchTriageRunner.java - Symptom / hypothesis:
JarConfigScanneronly checks whether the ZIP entry path for a declaredmain:class exists (mainClassPresent) and does not verify whether.classentries' compiledthis_classconstant-pool headers match their ZIP entry paths or whether embeddedMETA-INF/maven/<groupId>/<artifactId>/pom.propertiesmetadata conflicts withplugin.yml/paper-plugin.yml/fabric.mod.jsonidentity. - Impact: If an uploader renames a package directory inside a
.jarand updatesplugin.ymlmain:to match the renamed folder without recompiling the.classfiles,mainClassPresentpasses even though the JVM throwsNoClassDefFoundError: ... (wrong name: ...)on startup; uncompiled YAML rebrands over another author's build are also not surfaced as a dedicated triage finding. - Suggested next step: Investigate three static signals in
:bytecode-apiandLeafFuzzBatchTriageRunner: (1) deterministic ZIP entry path vs.ParsedClass#binaryClassName()(this_class) mismatch check (0%false-positive rate, catcheswrong nameNoClassDefFoundError), (2)META-INF/maven/.../pom.properties(groupId/artifactId) vs. descriptorname/author/mainpackage mismatch, and (3) normalized structural bytecode fingerprinting (class hierarchy + field descriptor multiset + method opcode skeletons) for recompiled clones. - Linear: RTP-104
2026-11-07 - evaluate_causal_rules_on_sources omits bodyless SPI interface methods from causal chains¶
- Severity: Low
- Status: Open
- Discovered during: running
scripts/tests/test_repo_ast.py(test_causal_chaining_scanner_and_declarative_rules) - Location:
scripts/repo_ast.pyevaluate_causal_rules_on_sources(~12924-13015);corpus-atlas/build_doc_index.pyextract_java_method_shorthand_cards - Symptom / hypothesis:
extract_java_method_shorthand_cardsonly extracts method scopes with{ ... }bodies, so abstract SPI interface declarations (byte[] loadSector(String path);inSectorLoaderSPI) are not added tomethod_nodesinevaluate_causal_rules_on_sources. As a result,CAUSAL-MAIN-IOresolves the 2-hop chain['TeleportController#onCommand', 'DiskSectorLoader#loadSector']viashort_method_lookupinstead of the 3-hop SPI bridge chain['TeleportController#onCommand', 'SectorLoaderSPI#loadSector', 'DiskSectorLoader#loadSector']. - Impact: Causal chains across pure interfaces omit the intermediate
SPI#methodhop intest_causal_chaining_scanner_and_declarative_rules. - Suggested next step: Synthesize stub
method_nodesentries for bodylessinterfacemethod declarations inevaluate_causal_rules_on_sources(or_extract_java_method_scopes) so theClosed-World SPI Dynamic Dispatch BridgeconnectsCaller -> SPI#method -> Impl#method.
2026-11-13 - compare_leafgen_vanilla_similarity.py skips :resource:LeafGen:leafgen-plugin:shadowJar when an older jar exists¶
- Severity: Low
- Status: Open
- Discovered during: live
LeafGenPaper26.2similarity scorecard verification (scripts/compare_leafgen_vanilla_similarity.py --run-axis0) - Location:
scripts/compare_leafgen_vanilla_similarity.pyensure_local_java_artifacts(line 250) - Symptom / hypothesis:
ensure_local_java_artifactsonly appends:resource:LeafGen:leafgen-plugin:shadowJarwhenleafgen_jar.is_file()isFalse, without comparing the existing jar's mtime against:bytecode-apior:resource:LeafGen:leafgen-commonsources/classes. - Impact: Running
compare_leafgen_vanilla_similarity.pyafter editingLeafGenJava sources without first running:resource:LeafGen:leafgen-plugin:shadowJarmeasures the previously built jar (deployed_sha == local_jar_sha). - Suggested next step: Always include
:resource:LeafGen:leafgen-plugin:shadowJar(Gradle incremental up-to-date check is< 1s) unless--skip-deployor--offlineis passed.