Log — FOSS Applications Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title>.

[2026-08-09] split | Spun out of the hub _inbox — six sources, five tags, nine weeks

The curator asked for a scan of the holding pen ("scan the inbox for clusters"), then approved this spin-out as a single spoke ("A and d"). The cluster was not visible in the records' own cluster: tags — the seven candidates were filed under five different tentative tags (local-first-foss-tools ×2, self-hosted-homelab, open-source-alternatives ×2, linux-desktop, windows-game-compat), so no tag ever reached the ≥3 trigger. Regrouping by actual subject made it the largest cluster in the inbox by a distance. Every record carried the same park reason in different words: no spoke owns software an end user runs themselves.

Founding corpus: six sources, all T1 primary repositories.

Concepts written: foss-application (umbrella), self-hosting, open-source-alternative. Entity: f-droid, paged because it recurs across two sources. Valve, Menny Even-Danan, yt-dlp and FFmpeg deferred at one mention each, per ../ENTITIES.md.

The seventh source did not come. termux was in the cluster and is already paged in ../osint-wiki as an OSINT enabler. Per the one-canonical-node rule it is linked cross-wiki as a bridge node and not duplicated here; that spoke's page was refreshed the same day with the repository's own numbers instead. The cluster is six, not seven, and it clears the trigger anyway.

A source contradicted the record that routed it. omniwm's park record — built from a showcase post on an aggregator site, with a broken canonical URL (https://github.com/ # OmniWM) — describes a C and Lua tiling manager for X11 with under 100 stars. The repository that same showcase names is Swift, macOS-only, GPL-2.0, 2.4k stars. The repository is what this spoke describes; both claims are on the page, and the aggregator is recorded as unreliable for this subject. Seven weeks of carrying a park record nobody re-checked against its primary source.

Thesis at spin-out (see synthesis): the six agree on the posture — the operator is in control — and disagree completely on its price. The finding that survives all six is that self-hosting relocates a dependency and is sold as deleting one, with hometube stating both halves in its own README. The second is that the category's own map (openalternative, star-ranked) cannot see compatibility work at all.

[2026-08-10] ingest | best-of — a directory of 22 generated, quality-scored open-source lists

Routed here by the hub (runner-up: ../engineering-education-wiki, which owns awesome-lists as curricula — this one teaches nothing, it ranks; ../dev-tooling-wiki declined for the generator alone under the standing Ky precedent). Source: https://github.com/best-of-lists/best-of, T1, CC0-1.0, 1,837★/150 forks read 2026-08-10.

Pages: best-of-lists (source summary), project-quality-score (the metric), best-of-generator (the tool, thin node). openalternative gained a cross-link; synthesis gained a second half to its "the category's own map cannot see part of itself" section.

Why it earned three pages rather than one. The spoke's third finding rested on a single directory's ranking method — openalternative sorts by stars alone — and had no counter-example. This is the counter-example: the same genre, ranking by a score that gives one point each for a licence, semver, a release inside six months, a commit inside three, and adds stars as round(log(n)/2) − 1 among a dozen logarithmic terms. Read from the generator's source, not its marketing.

What it does not fix. Both maps rank repository metadata. Neither reads code, runs the software, or counts users — and the one usage term the score does have, monthly downloads, is unearnable by software that ships as an application rather than a package. The generator's docs concede it outright: "There is no scientific proof that this really reflects the quality of a project."

Finding from the source's own markers. 17 of the 22 indexed lists carry 💤 (six months idle) or 💀 (twelve months idle). The apparatus built to detect abandonment is mostly documenting its own — a generated list keeps rebuilding after the curator stops choosing what goes in it.

Entities: Benjamin Räthlein, Lukas Masuch, Jan Kalkan and Johannes Rieke named once each as best-of-generator maintainers; deferred per ../ENTITIES.md, as Valve and Menny Even-Danan were.

[2026-08-10] ingest | OpenPanel — AGPL analytics that calls itself the open-source Mixpanel

Routed here by the hub (runner-up: ../analytical-databases-wiki, which owns the ClickHouse question and already holds posthog-duckdb-vs-clickhouse — the same product category making the same engine choice; ../search-marketing-wiki was the second runner-up on the measurement angle). Source: https://github.com/Openpanel-dev/openpanel, T1, AGPL-3.0, 6,449★/418 forks/77 open issues, read 2026-08-10.

Pages: openpanel (source summary). self-hosting and open-source-alternative both updated; synthesis gained a third finding and a sixth open question.

The finding. Analytics inverts the direction every privacy claim in this corpus had pointed. Until now the beneficiary of self-hosting was the operator (anysoftkeyboard cannot leak what it never sends). Here the people recorded are the site's visitors, and moving the software in-house does not reduce the recording — it removes the vendor policy, the audit and the data-processing agreement that stood between them and the data. Relocation, again, but this time the risk lands on someone who never made the decision.

Two secondary gains. Second AGPL-3.0 project here selling its own hosted version after ever-gauzy, so open question 5 has a pattern. And its docs state the recurring maintenance outright — updates mean reading the changelog and applying changes to your instance — which is the first direct evidence for open question 2.

On the imitation hypothesis. This is the only source that applies the open-source-alternative label to itself, in its own GitHub description, and its feature list is Mixpanel's. Recorded as one ambiguous instance, not as a confirmation — those features are also what anyone shopping for analytics asks for, and nothing separates the causes.

Read from the repository, not the marketing. The seven-container stack comes from self-hosting/docker-compose.template.yml; the docs page describes the stack loosely and never enumerates it.

[2026-08-13] ingest | PRoot-Distro — rootless Linux containers on Android

Routed here by the hub from a Telegram link (github.com/termux/proot-distro). Runner-up: ../osint-wiki, which pages termux in full as an OSINT enabler — but this is the environment, not the recon kit, and this spoke's boundary rule takes software an end user runs on a machine they control. ../dev-tooling-wiki was considered and declined on its own Ky precedent: a runtime a person computes in is not build-and-ship infrastructure.

T1, README and limitations documentation read in full plus the GitHub API for the counts. GPL-3.0, Python 3.9+, 3,380★ / 423 forks / 5 open issues, created 2020-07-20, pushed 2026-08-13 — the day it was read, so the numbers are a same-day snapshot of a moving project.

New pages: 2. proot-distro and compatibility-layer (a concept the spoke was already carrying with one source and can now hold with two). Updated: synthesis, index.md.

What the tool is now, versus what it was. The 2020 version shipped its own prebuilt rootfs tarballs. This one is an OCI registry client: it speaks the Distribution protocol over urllib, takes an anonymous bearer token, resolves manifest lists to the local CPU, verifies each layer's SHA-256, applies whiteouts, and caches so a reinstall works offline. It also builds images from a Dockerfile with no daemon and pushes them back. The distribution channel moved from the project's own files to Docker Hub, which is worth noticing in a project whose premise is not needing permission.

The register finding. With proton-wine-andrevto this makes two sources whose subject is neither an environment substitute nor a self-hosted service: both make software run where its author never targeted it, and both do it in user space — Wine reimplements Win32, proot intercepts syscalls via ptrace. compatibility-layer carries the comparison and the reason openalternative's taxonomy cannot see either.

Relocation sold as removal, in the code this time. It removes root, the bootloader unlock, the vendor and the daemon — then replaces every container's /etc/resolv.conf with 8.8.8.8 / 8.8.4.4 and defaults its images to Docker Hub. Documented plainly in the README, presented as a fixup rather than a choice. Also recorded: PD_DOCKER_AUTH=user:password is the only credential path, since Docker credential helpers and ~/.docker/config.json are deliberately unread.

Best-documented failure surface in the corpus. ptrace overhead on every syscall, no real root, no kernel modules or FUSE, no init system, no cgroups or namespaces, no nesting, no zstd layers, and BuildKit-only Dockerfile directives rejected with an explicit error. The spoke's standing caveat is that claims here are self-reported; this one publishes its own failure list, which is not verification but is the first source that makes verification easy.

Reflexive: the hub runs under PRoot-Distro (uname -a → 6.17.0-PRoot-Distro … aarch64), so the costs on this page are the ones every build in this repository pays.

Entities: 0 created. The Termux project is already paged as termux in ../osint-wiki and bridged from here; no author entity is evidenced beyond the org.

[2026-08-17] ingest | TUIOS (Gaurav-Gosain/tuios)

Arrived via Telegram, routed by the hub (runner-up: ../ui-frameworks-wiki). Routed here on this spoke's own written boundary — the toolkit an application is built with is there; the application is here — so the Charm stack (Bubble Tea v2, Lipgloss v2) is that spoke's subject and cross-linked, and the multiplexer is ours. That spoke holds terminal-ui and opentui and has no Charm page yet.

tuios, T1 (primary repository, figures read from the GitHub API rather than the README): MIT, Go, 3,250★ / 115 forks / 20 open items / 9 watchers, 1,161 commits across 12 contributors with 1,101 by the author. Created 2025-09-06, last pushed 2026-08-16 — active. Latest tag is v0.7.0 from 2026-03-28, so main has run four and a half months past the release; Homebrew and go install @latest are handing out different programs, which is on the page as an unexplained gap.

The finding is the pairing with omniwm. Two single-author tiling window managers, same tradition, same ambition, and not comparable at all. OmniWM fights macOS private APIs, pins to macOS 26+ on Apple Silicon, and pays Apple's notarization toll to be installable. TUIOS does the job inside a terminal emulator, a rectangle the user's process already owns — nothing to ask for, nothing to intercept, no update that can revoke it. That is a fourth control position, cheaper than compatibility-layer's reimplement-or-intercept, and it reframes the spoke's recurring complaint: the friction these projects meet belongs to locked platforms, not to independent software as such. The ceiling is stated too — it manages terminal panes, not applications.

Distribution says the same thing from the other end: six channels (Homebrew, AUR, Nix, ghcr.io, go install, shell script) for a one-person project, because nobody polices terminal programs, against anysoftkeyboard and termux needing f-droid to exist at all.

Also the spoke's first control case on motive — the README argues about none of the things every other source here argues about (clouds, vendors, privacy, data) and pitches performance engineering instead. Same posture, no ideology. Recorded weaknesses: no benchmark behind any of the performance numbers (incl. the claimed 40–60% allocation reduction), no comparison to tmux/Zellij/WezTerm, bus factor of one, and several graphics features the project itself calls experimental or broken. synthesis.md and index.md updated.

[2026-08-17] ingest | awesome-terminals (cdleon)

Arrived via Telegram, routed by the hub (runner-up: ../ui-frameworks-wiki, which owns terminal-ui as a toolkit layer — an emulator is software an end user installs, and a catalogue of them is this spoke's third register). Counted from the raw README rather than taken from a summary: 403 entries — 129 emulators across seven platform sections (Linux 48, macOS 37, Windows 26, iOS 7, Web 6, Android 4, HarmonyOS 1), 253 CLI tools by implementation language (Go 55, Rust 48, Python 39, Node 28, C 23), 8 shells, 13 dotfile repos. 2,925★/159 forks/19 open items/267 commits, created 2018-10-07, pushed 2026-07-20. No licence file. T3 — no inclusion criteria, no dates, no metric, nothing verified.

The finding reframes the spoke's second one. openalternative ranks substitutes by stars and cannot see proton-wine-andrevto or omniwm; best-of-lists answered with a better metric and changed nothing, because both rank the same thing. This list changes the axis — platform and language, never "what does this replace" — so that blind spot cannot occur in it. The problem was the substitution frame, not star-ranking. The new blind spot is the mirror image: a category axis admits the whole category, so iTerm2 and a dead commercial product are both in here and this is not a FOSS directory. Recorded in synthesis.md as a correction to the second finding rather than a new one.

Two smaller things kept. Liveness is marked by a human, thirteen hand-written [DEPRECATED] tags, against best-of-lists's automated 💤/💀 — a judgement that can know an author declared a project finished, where a threshold is consistent and blind; the corpus now holds both. And the list declines to rank at all, which is worth setting against project-quality-score's premise that what the field needed was a better number.

Also flagged: 403 curated entries with no licence, where both sibling catalogues are CC0-1.0. For a directory the data is the value, so the missing permission bites harder than it would on a tool. tuios, ingested here the same day, is not in the list — it is a multiplexer, not an emulator, which is the boundary this catalogue draws by construction.

[2026-08-22] ingest | 2025 Self-Host User Survey (via quality cycle)

self-host-user-survey-2025 — selfh.st's annual reader survey, 4,081 completed responses, published 2025-11-21. T2: an independent publisher surveying users of software it does not make, and the load-bearing claims are its own collected counts. Figures read from the project's published JSON (selfhst/cdn, assets/surveys/annual/2025-results.json) rather than from the charts, so every number on the page is countable.

Ingested against growth edge 2 ("a survey or dataset on self-hosting practice — who runs what, for how long, why they stop"). It is a half-close: the survey asks what people run now. There is no tenure question, no churn question and no abandonment question, so open question 2 — what happens when a maintainer stops — gets nothing at all. Edge 2 keeps its retention half and a new edge 2a records the second residue: the sample is one newsletter's audience, 81% in technology and 94% male, so the spoke still has no evidence about non-enthusiast operators.

What it did change: open question 3 moves to partly answered, and synthesis.md gains a finding. The motive ranking is the surprise — hobby 3,862 ahead of privacy 3,520 and cost 2,596 — which sits against a corpus read entirely off privacy and independence claims in READMEs. The second is that 85% update containers manually or case-by-case, 77% run no orchestration, and 34% keep a single copy of their data: the maintenance burden self-hosting calls the hidden price is visible to these operators and accepted rather than automated away.

Also recorded: 8 enterprise respondents out of 4,081, which is why the survey cannot be used as evidence about ever-gauzy-class business adoption; and Proxmox (1,832) plus Home Assistant OS (1,199) outranking Raspberry Pi OS (979), the community moving toward prebuilt appliance substrates.

No dedup collision — the only corpus hit for "selfh" was a string inside defensive-security-wiki's vaultls page ("Selfhostable web app…"), unrelated.

[2026-08-25] ingest | Termux Launcher — an Android home screen that is a terminal

Arrived by Telegram as a bare GitHub URL and routed here by the hub. Read the repository README in full plus the GitHub API metadata; T1, first-party. 190★ / 14 forks / 2 open issues, created 2026-01-30, last pushed the same morning it was read. New page: termux-launcher. No existing page covered it — grepped every spoke for termux-launcher, termux-monet, nix-on-droid and nixpkgs first, and termux itself stays where it is, paged in ../osint-wiki and linked as a bridge node.

Three things this source adds that the corpus did not have.

First, a fifth control position. Every earlier entry takes, works around, reimplements or avoids the platform. Android simply hands over the home-screen slot to whoever the user picks, so a terminal becomes the phone's main surface through a documented extension point. Cheapest position held here, most invasive result. Written up in synthesis.

Second, distribution by taking the upstream's package name. The recommended edition ships as com.termux and therefore removes official Termux from the device, while the README disclaims any affiliation in bold and asks users not to file bugs upstream. Set beside f-droid as an alternative store, tuios's six package managers, and proton-wine-andrevto's repackaging, that is a fourth and sharper answer to how independent software reaches a user.

Third, the first evidence for open question 2 — what happens when a maintainer stops. The VAJ edition is being deprecated because hand-maintaining an APT repository for security updates is "no longer sustainable," and a migration guide to the Nix edition comes with it. Supply side, one edition, self-reported. The question stays open; it is no longer empty.

Recorded as unverified: the "rock-solid" claim is one author on one Nothing Phone (2); the signing-key inference behind the add-on pairing rule is mine, not the README's; and the AI-assisted app categorization has no stated execution location. The author's own header — "This project is entirely vibe-coded" — is on the page as an assurance gap rather than a defect, and is the first maintainer warning in this spoke about how code was written rather than what it does.

Cross-spoke, not duplicated: on-device inference (LiteRT, MNN, an OpenAI/Ollama-compatible localhost API) belongs to ../llm-inference-wiki (local-llm-stack) and phone-side agents to ../agentic-tooling-wiki (hermes-agent-termux). Both linked from the page and from synthesis → Cross-spoke adjacency.

Entities: none paged. PickleHik3, Termux:Monet, TEL, LiteRT, MNN, Shizuku and Nix/nixpkgs are one mention each and deferred per ../ENTITIES.md; noted in index.md.

[2026-08-25] ingest | TEL — the launcher Termux Launcher was copying, dead since 2021

Second Telegram source of the day, sent about an hour after termux-launcher and pointing at the project that one names as its inspiration. Read the repository, LICENSE.md, .travis.yml, the release list, the contributor list and the project wiki at t-e-l.github.io/wiki (eight sections; the applet and status-manager pages carry the substance). T1, first-party throughout. New page: tel-termux-expert-launcher.

154★ / 8 forks / 6 open issues, Java, created 2020-03-05, GPLv3 over Apache-2.0 Terminal Emulator for Android code. Last master commit 2021-11-25; last release the same day, 9,690 downloads. Never archived. Contributor list shows fornwall at 501 commits — inherited termux-app history — with v1nc (296) and SealedJoy (181) as TEL's own authors.

How it works, in three parts: a [TEL] leader key in the extra-keys bar over tmux (USE_TMUX=true is required, so panes and windows are tmux's), fourteen applets that replace phone apps (tel-dialer calls contacts through termux-api, aliased to p), and a status bar that is a directory — any .sh/.py printing to stdout is a module, disabled by moving the file to status/inactive.

Two admissions taken from the docs and kept on the page. The status manager "is known to potentially cause high battery usage," with a whole FAQ entry on the drain. And input-char in settings.conf exists because the app search reads its own terminal output: change your shell prompt and the home screen stops working. Both are the project's words, not an outside audit.

Three findings, written up in synthesis.

1. Open question 2 answered from both sides in one day, and only from the supply side. termux-launcher retires an edition on purpose with a migration guide; TEL's maintainers stopped with no notice and no successor, and the design was rebuilt five years later by a former user who wanted sixel graphics. Forkability preserved the possibility of continuity and delivered none — the ~9,700 downloads still hold a 2021 APK. Added growth edge 2b for the missing user-side account.

2. Killed by its build server, probably. Travis CI pipeline, APKs named from TRAVIS_BUILD_NUMBER, an encrypted keystore.jks.enc committed for CI to decrypt; Travis ended its free OSS tier in 2021 and the releases stop exactly there. Recorded as inference, since the commits stop too. It names a dependency the spoke had not examined — every other page asks what a project costs its user, this one lost distribution to a build service while its source stayed fine.

3. Forking Termux means forking its add-ons. TEL ships TEL:API and says only that build may be used; termux-launcher ships its own Termux:API and Termux:Styling because F-Droid's "will not pair with any edition." Two forks, five years apart, no contact, identical constraint — Android's package-family-plus-signing-key identity model charges it. Yesterday this was an inference on one page; with a second independent case it is on synthesis as a rule.

Also recorded as unverified: the install base (9,690 is a download count), that nothing here has been run, an August 2024 push to a non-default branch that nothing explains, and a hole in the attribution — LICENSE.md reads "Contains code from Terminal Emulator for Android by which is released under the Apache License 2.0", with the author's name missing from the sentence.

Entities: none paged. TEL left the deferred list by being paged itself. v1nc, SealedJoy, Fredrik Fornwall and Travis CI are one mention each and stay deferred per ../ENTITIES.md; Fornwall's commit count here is fork history, not participation in TEL. termux-launcher updated to link TEL instead of naming it in bold.

[2026-08-29] ingest | calibre — About page

Routed here by the hub router from Telegram (runner-up: ../research-wiki). Model: claude-opus-5 · effort: high.

calibre (T1, the project's own About page, read 2026-08-29). An e-book manager for Windows/Mac/Linux started 31 October 2006 by Kovid Goyal, as libprs500 — reverse-engineered USB protocol work to make a SONY PRS-500 talk to Linux — renamed calibre in mid-2008. Claims more than 3 million active installs across 200+ countries, a dozen volunteer translations, hundreds of one-click plugins, and donation funding (Patreon + PayPal). Eight capability groups: library management, conversion across dozens of formats, device sync, viewer, RSS/news delivery, content server, EPUB/Kindle editor, plugin explorer.

Three firsts for this spoke. It is the oldest project here by fourteen years. It is the first source that is a project's self-description rather than a repository — no commit count, no forks, no issue tracker, and no licence field. And it is the first install figure the corpus holds, which is the measure project-quality-score explicitly cannot compute for software that ships as an app.

Two defects recorded rather than smoothed. The page contradicts itself on development size — "hundreds of motivated volunteers" in the feature block against "half a dozen developers" in Goyal's own account. And half the page is a memoir signed "Kovid Goyal, November 2009" sitting unrevised under present-day statistics, so published: was deliberately left off: dating the document to 2009 would misdate the numbers, dating it to today would misdate the essay. The licence is never named on the page; recorded as GPL-3.0-only from Wikipedia (T3, checked 2026-08-29) and flagged as wanting the project's own licence file.

Updated: foss-application (a project in three registers at once; the install figure against "no evidence of use"), project-quality-score (calibre as the case that shows the metric's ceiling), compatibility-layer (a third case, about content rather than programs — held as an adjacency, not folded in), self-hosting (the content server as the corpus's smallest self-hosted thing), index.md. Synthesis: new section "Twenty years is a different kind of evidence"; open question 3 gains a supply-side number; new open question 7 (does anything here survive its founder?); growth edges 5 and 6 added.

Entities deferred under this spoke's recurrence rule: Kovid Goyal, Krittika Goyal, mobileread.com, Marshall Vandegrift — all named by this one source. Goyal is the strongest deferral candidate the spoke has had; a second source pages him.

[2026-08-30] ingest | Alexandrie — self-hosted wiki and knowledge base

Routed here by the hub router from Telegram (runner-up: ../research-wiki). Model: claude-opus-5 · effort: high. Read from the repository README, the GitHub API, and docker-compose.yml.

alexandrie (T1). MIT, by Smaug6739, 2,633★ / 179 forks, Vue/Nuxt frontend + Go/Gin backend, created 2022-08-21, pushed 2026-08-29. A self-hosted wiki with an extended CodeMirror 6 markdown editor (LaTeX, containers, inline diagrams, snippets, dictation), a multi-tenant Teams→Workspaces→Categories→Documents hierarchy with a 5-level permission model, OIDC/SSO, Kanban boards, Ctrl+K command centre, full-text search, S3 media storage, ZIP import/export, and an offline-first PWA. docker compose up -d, owner account at localhost:8200.

Why it earns more than a catalogue line. Its own GitHub description is the purest instance of the open-source-alternative frame this corpus holds: "The open-source, offline-first Notion, Obsidian & Confluence alternative" — three proprietary products named before the software says what it is. That makes it the second case for the spoke's open question 4 (does the frame change what gets built, or only how it is sold?), and a much less ambiguous one than openpanel, which named a single target.

The finding, and it is new. The imitation is legible feature by feature — Notion supplies nested documents, Kanban and the command palette; Confluence the teams, permissions and SSO; Obsidian the markdown and offline mode. But Obsidian's actual proposition is a folder of markdown files on your own disk, and Alexandrie stores documents in MySQL, offering a ZIP export as portability. So the frame carried the nameable feature and dropped the architectural one. The substitution frame transmits descriptions, not designs — which predicts alternatives that match a product's bullet points while missing the property its users chose it for. Written up in synthesis.md as "The frame copies the description, not the design."

Operational cost, read from the compose file because the README does not state it. Four services, two stateful: mysql:8.0 and rustfs (S3-compatible), plus the Go backend and the Nuxt frontend. That is two backup jobs the operator now owns, on a project whose README promises "total ownership over your binaries, media streams, and backups" and never mentions a database. Slots into self-hosting's spectrum just below openpanel (7 containers, 3 stateful) and well above hometube (1).

Also recorded as absent: no requirements (RAM/disk/CPU) for a stack containing MySQL and an object store; no security or audit history on software shipping SSO and five permission levels; no comparison against any of the three products it names.

Smaug6739 not paged — a publishing handle with no biography evidenced, following the AndreVto and PickleHik3 precedents in this spoke.

New page: alexandrie. Updated: open-source-alternative (the clearest imitation case, and where imitation stops), self-hosting (fails the name-the-resource test from the inside — the unowned resource is the backup job you were not told about), foss-application (register 2), index.md, synthesis.md (new section; open question 4 narrowed rather than answered).

[2026-09-01] ingest | Termux Launcher, re-read a week on

The curator re-sent github.com/PickleHik3/termux-launcher, already held since 2026-08-25. Refreshed termux-launcher in place rather than writing anything new; both readings are kept on the page.

Growth. 190★ → 239★, 14 → 17 forks, 2 → 3 open issues, 6 watchers, last push 2026-08-31. Fifty stars in a week for a project distributed as bare APKs from GitHub Releases with no store listing.

The stability claim changed shape. The README's warning box said "rock-solid so far" on 08-25 and now says "stable so far" plus two numbers: no noticeable battery impact, and ~350 MB of RAM at idle against ~250 MB for Smart Launcher on a fresh install. Weaker adjective, harder claim — the opposite of the direction first-party copy usually travels, and still one device with no method stated.

The finding worth the re-read: a deprecation with a tail. Release v0.2.38+hotfix1 shipped 2026-08-29 in all three editions at the same minute, and three days later the assets read 126 downloads for the recommended Termux edition, 75 for the deprecated VAJ edition, 23 for the Nix edition meant to replace it. The retiring build is being taken three times as often as its successor, and the maintainer is still hotfixing it. For the spoke's open question 2 that sharpens the question: the cost of a maintainer stopping is not concentrated at the moment they announce it but in the interval where they have announced it and are still carrying both. Recorded with its limits — downloads on one release over three days, not installs and not a trend.

No new pages, no new entities. check:wikilinks and check:arguments clean.

[2026-09-01] ingest | PocketPal AI — language models on the phone, and the store it ships through

github.com/a-ghorbani/pocketpal-ai — MIT, React Native/TypeScript, iOS + Android. Routed here by the hub; runner-up was ../llm-inference-wiki, which owns the engine (llama-cpp, local-llm-stack) while the installable application is ours — the same call the spoke made for termux-launcher. Read from the README and the GitHub API on 2026-09-01: 8,137★ / 850 forks / 181 open issues / 67 watchers, created 2024-08-25, 741 commits, v1.17.2 (2026-08-25). T1, the project's own repository. New page pocketpal-ai.

Two findings, and the second is the one the corpus did not have.

The privacy claim comes with its own subtraction. "Every prompt, response, and document stays on your device" is followed immediately by a privacy note listing what does leave: opt-in benchmark submissions and in-app feedback. Model downloads, Hugging Face tokens for gated models, and the PalsHub marketplace (with in-app checkout for premium personas) are legible elsewhere in the same README. That inverts hometube's shape, where the reader has to find the contradiction. Recorded with the sting: this is simultaneously the corpus's most honest privacy claim and its least checkable one, because unlike anysoftkeyboard the app holds the network capability the promise asks you to trust it not to misuse. Open question 1 updated — the question is now whether a claim of this shape can be checked from outside at all, and growth edge 1 now asks for a network capture rather than a permissions read.

Refusing the cloud and choosing F-Droid are separable — and a channel can be real and unadvertised. The README offers the App Store and Google Play and nothing else, and there is no f-droid entry (the F-Droid API 404s for com.pocketpalai, checked 2026-09-01). The first draft of this ingest said there was no APK either, and that was wrong: every release attaches app-prod-release.apk, and the counts are large — 6,154 downloads on v1.17.2 within a week of release, 3,520 on v1.17.1, 3,144 on v1.17.0, 16,891 on v1.16.1. Corrected before the page was committed, and the correction is the better finding: a distribution channel can carry thousands of installs while appearing in no documentation at all, which means counting channels from a README undercounts them. New synthesis section "Two autonomies, and they come apart", and the f-droid node gained a counter-case section — paged to describe what a project pays for leaving the store, it now also holds the case of one that never put up a sign.

Also recorded: the four-layer stack (llama.rn → llama.cpp/GGUF, ONNX Runtime for TTS; CPU / GPU Metal-and-Adreno / NPU Qualcomm Hexagon with fallback), an on-device tool-use loop with three "Talents" and expanding it as the only roadmap item, 11 locales on Weblate, and a project-run AI Phone Leaderboard fed by opt-in submissions — the first product here that ships its own measurement apparatus, and user-submitted data is what it is worth. Contributor concentration is the sharpest yet: 494 of 741 commits the author's, with a bot second at 104.

No new entity pages. Asghar Ghorbani and six others are named by this one source and deferred per ../ENTITIES.md; a name collision with the Exchange Market Algorithm's Ghorbani in ../optimization-algorithms-wiki is recorded on the page and in index.md so a later ingest does not merge two people.

[2026-09-01] ingest | Ceefax Station — teletext over amateur radio

Routed from the hub (runner-up: ../embedded-iot-wiki, on the Raspberry Pi / radio-protocol angle; the subject is an application an end user installs, so it landed here).

ceefax-station (T1, the repository) — the BBC's Ceefax service as a curses viewer, with an optional AX.25 AFSK1200 transport so licensed operators can send and receive pages over 2m. Python 3.11, MIT, 9 stars, 0 forks, v0.1.5-alpha, created 2025-12-29, last push 2026-08-30, .deb + Windows EXE with in-app self-update.

The finding is a split the corpus has not seen in this shape. The radio half has no server in it at all — packets go station to station — while every news, sport, lottery, TV and weather page is downloaded pre-built from ceefaxstation.com, a hub the same author runs. The README states this in a glossary and a provenance table and frames it as convenience: "You do not need API keys", because the hub holds them. The fallback sentence gives it away — with the hub down, the app builds pages locally and "some pages then need API keys." That is a third posture on the spoke's founding finding, next to hometube (omits) and pocketpal-ai (subtracts for you): fully disclosed, inverted framing.

Two more. It uploads the operator's callsign and Maidenhead grid to a public map by default after any TX/RX session — opt-out, documented, never called telemetry, and a callsign is a legal identity. And it is the first application here with no utility argument: nobody needs teletext, which makes it the corpus's first match for the hobby motive that self-host-user-survey-2025 found outranking privacy and cost.

New: ceefax-station. Updated: index.md, synthesis.md (new section — a third split, a third posture, the default upload, and the utility break). No new entity pages: thaum-labs / M7TJF is named by this one source and deferred under ../ENTITIES.md; nothing here verifies an identity beyond the callsign and the project domain.

Recorded as unknown: what happens to installed stations if the one author stops running the hub, the tracker and the domain.

[2026-09-04] ingest | Locally Uncensored — a local AI studio with a second mode

Model: claude-opus-5 · effort: high. Routed here by the hub from Telegram, one of three sources arriving together. locally-uncensored is an AGPL-3.0 Tauri/React desktop application bundling chat, a coding agent, and ComfyUI image/video generation behind one Windows or Linux installer — 1,374★ / 213 forks / 7 open issues / 1,317 commits, created 2026-03-24, v2.6.7, read 2026-09-04.

The finding is a new posture on the founding one. The corpus had three readings of relocation sold as removal, and all three were about disclosure: what the README says and where it says it (hometube buries, pocketpal-ai discloses adjacently, ceefax-station discloses fully and frames it away). This source's README is accurate on both counts — "Free forever… no credits" in the headline, LU Cloud with a credits meter and per-second music billing in the changelog — because they describe two modes of one application. What makes it new is the 2.6.4 fix, in the project's own words: flipping to Local with no local model running "used to keep the cloud model silently active and kept billing credits." The claim was enforced by a conditional rather than by an absence, the conditional was wrong, and no amount of reading the document would have protected the user. Synthesis gains "The fourth posture, and this one has a meter", and growth edge 1 is re-aimed at a teardown of a dual-mode application — the local/cloud boundary tested under the states a user can actually reach.

Two smaller keeps. The spoke's first Windows-first source, and the first whose audience is legible from the release page: 8,563 .exe downloads against 576 across every Linux artifact on v2.6.7, with latest.json at 11,308 behind them. And a third distribution position beside f-droid and termux-launcher — no store at all, but a signed auto-update channel the project runs itself, answering antivirus flags with a reproducible-build argument and its own published minisign key.

Cross-spoke, not duplicated: abliteration and heretic stay in ../llm-providers-wiki, which owns the mechanism and its measured cost. This page is the consumer end — the first source in the hub that turns an abliterated model into a hardware-filtered, one-click menu item for a non-technical Windows user. Deferred entities: PurpleDoubleD / "LU Labs", ComfyUI, Tauri, faster-whisper, Kizerfluid — one source each. Pages: 1 source + 1 synthesis section = 1 new. Runner-up spoke: ../llm-providers-wiki (the uncensored framing); routed here because the subject is an application an end user installs and runs.

[2026-09-07] ingest | Syncthing — peer-to-peer file sync, and the smallest dependency yet

Telegram source, routed here by the hub over ../cloud-wiki. syncthing — T1, the repository, with the front page and two documentation pages read alongside it.

The numbers. MPL-2.0, Go, 88,364★ / 5,455 forks / 381 open issues, repo created 2013-11-26, v2.1.3 released 2026-08-05. The largest and oldest project this spoke holds. Not new to the corpus, though: self-host-user-survey-2025 already had it second in Files with 1,058 selections, behind Nextcloud — so the page was written against a number the spoke already had, and that link now runs both ways.

Applied this spoke's own test and the answer is unusually small. Name the resource the software needs that you do not own. There is no host, no account and no storage anywhere but your machines. What remains is a rendezvous layer: global discovery resolves a device ID to an IP, and relays — "run by volunteers all over the world", enabled by default, used only when two devices cannot connect directly — carry the encrypted stream. Quoted from the relaying docs: the relay "only retransmits the encrypted data much like a router", it "knows your IP and device ID", and the operator "can see the amount of traffic flowing between devices."

That gave synthesis a new section, "The dependency shrinks to a rendezvous." The corpus had been treating the dependency as one axis, present or removed. It has two: what the third party holds (files vs metadata) and when it is involved (always vs on failure). This is the first source scoring well on both and non-zero on either.

Two smaller findings. The README — 470 words, a ranked goal list with data integrity above security above usability — makes no privacy claim and never says "cloud", which makes the strongest instance of this spoke's subject also the one that markets itself least. And the disclosure failure is the ordinary kind: the loose sentence ("there is no central server that might be compromised") is on the marketing page while the manual spells the dependency out in more detail than a user would ask for. That is the inverse of hometube, which buried its cookie requirement, and the two should not be scored the same.

New entity: kastelo (Kastelo Inc.), which sells commercial support for Syncthing and sponsors development while the project stays MPL-2.0 and community-run. A third answer to who is paid to notice when it breaks, beside ever-gauzy's nobody — and unquantified, because the sponsored project is the one describing it.

Not verified: the encryption claims, the relay behaviour, and whether global discovery can be switched off — the FAQ passage read here does not say and the configuration reference was not opened.

[2026-09-10] ingest | Olauncher — Minimal AF Launcher

Routed in from the hub (Telegram). Read the GitHub API for the figures and the master README (the default branch is not main — main returns 404).

Source: olauncher, T1, the project's own repository. 3,803★ / 496 forks / 25 open issues, created 2020-07-10, last pushed 2026-09-01, Kotlin, GPL-3.0, on F-Droid and the Play Store. published: 2020-07-10 (repository creation, the only date the source states).

What it changes: the spoke held two Android launchers and both were terminals for experts (termux-launcher 239★, tel-termux-expert-launcher 154★). This is 15× the larger of them and a different species — keep the GUI, remove nearly everything in it, for screen-time reduction rather than for shell access. Synthesis now separates the two minimalisms instead of treating "minimal launcher" as one category.

Two smaller findings recorded: features deliberately hidden to protect the default user, which inverts this spoke's usual treatment of configurability as the point; and a clean small-scale open-core specimen — GPLv3 app, closed paid Pro Launcher from the same author with no repository, plus donations for a test phone.

Bears on growth edge 6 (single-maintainer governance and succession): six years, 3,803 stars, no governance document, and a commercial product built on the free app's users. It is the best-documented instance of the risk the edge names and not an answer to it, which is how the page states it.

Not read: any privacy or permissions claim (none made), release cadence, contributor counts, and the in-app About page where the feature list actually lives. The README is unusually thin.

[2026-09-27] ingest | Quickemu

Routed in from the hub with high confidence, no runner-up spoke. Read the GitHub README and the wiki's table of contents (not the individual wiki pages — narrow reading), plus the GitHub API for repository metadata.

Source: quickemu, T1, the project's own repository. MIT, 16,330★ / 744 forks / 72 open issues+PRs, created 2020-03-15, pushed 2026-09-14, org quickemu-project. published: 2020-03-15 (repository creation, the only date the source states). Two shell scripts around QEMU/KVM: quickget fetches an OS image and writes a config, quickemu reads the host's hardware and boots a tuned VM with no elevated permissions. Claims nearly 1,000 supported OS editions across macOS, Windows, Ubuntu, most BSDs, and niche systems (FreeDOS, Haiku, KolibriOS, OpenIndiana, ReactOS); hosts are Linux and macOS only.

Gap-relevance: tangential to the named open questions and growth edges (none is about virtualization), but a direct hit on the spoke's domain and a genuine new mechanism. rg'd the corpus first — nothing already held on QEMU, KVM, or hypervisors.

What it changes: compatibility-layer held two sources on one mechanism — reimplementing (Wine/proton-wine-andrevto) or intercepting (proot-distro) an interface so foreign software runs somewhere its author never targeted it. Quickemu reaches the same destination by the opposite route: it fakes nothing and boots the guest's own unmodified kernel under hardware virtualization. Recorded as an adjacency, not folded into the register — same treatment calibre's device- protocol history got on 2026-08-29 — with a note on both pages and a new dated section in synthesis.md.

One finding sharpens the corpus's standing permission question. Quickemu tags itself hackintosh and lists macOS as a guest, which is exactly the use Apple's own licence restricts to Apple hardware. omniwm pays Apple's toll for the equivalent privilege (notarized, signed builds); Quickemu does not ask and does not pay — the licence becomes the host operator's problem, unmentioned in the README. Read alongside proton-wine-andrevto (reimplement around the vendor) and omniwm (pay the vendor), this is a third posture: take the real hardware and skip the question.

Entity discovery: GitHub's org social links point to Martin Wimpress (flexiondotorg); project wiki and podcast notes credit the Linux Matters hosts with a 2024 revival. Both named by this one source — deferred per the spoke's recurrence convention, noted in index.md → Notes, same treatment calibre's Kovid Goyal got despite being the more documented case. No entity pages created; budget not spent.

Not read: the ten-page GitHub wiki past its table of contents (installation steps, per-platform guides, governance if any live there), and no independent check of the "nearly 1,000 editions" claim or of any of the pass-through/acceleration features working as described.

[2026-10-11] lint | absence-claim drain (step 3a), 2 claims on proton-wine-andrevto

"The only source in the corpus that runs someone else's proprietary software" was retracted in place: quickemu (2026-09-27) boots Windows and macOS guests. "No test evidence of any kind" still holds, because the repo README has no compatibility list.