Engineering evolution
The history of the knowledge system itself: what was learned, built, decided, and published — and when. Not a résumé.
The Lab shipped as a collection of living configurations plus a generated experiments collection. Configurations separate machine-readable inventory (companion metadata) from human documentation; experiments reference a configuration through a runs-on edge. First consumer of the media blocks.
Compass became true navigation memory. It now appends on every navigation — including revisits, so a loop like A → B → A is recorded rather than collapsed — and truncates only on a deliberate rewind, when the reader clicks a node in the Compass. This supersedes the v1 rule that truncated on any revisit (ledger ADR-004); v1 is archived, not deleted.
Citations became a single authored form: the inline :cite[id]{at=…} directive in prose. From it the build derives everything a reader sees — the numbered inline marker, the document's References section (catalogue metadata plus the instance's locator), and the cites edges in the knowledge graph (previously hand-authored). Free-text references frontmatter is gone; to cite is to catalogue (ledger ADR-005).
The contact form gained a real submission path: a single dynamic endpoint (/api/contact) that validates, screens for spam (honeypot, submission timing, per-IP rate limiting), and writes one markdown-with-frontmatter file per message to a private store outside content/. It is the sole non-static route — a write-only inbound ingress that never becomes knowledge — so the static-output invariant on the knowledge system is preserved (ledger ADR-006). Persistence sits behind a swappable seam so the production sink is chosen at integration.
The first integration-phase pass over the finished architecture. Section identity — the label, route, nav position, hue, and chip of each section — was scattered across ~8 files in three incompatible string conventions, where a plural/singular slip dropped a hue with no build error; it now lives in one typed registry that everything derives from (ledger ADR-007). The notebook page began honouring its spec'd last-touched date (distinct from updated); the duplicated experiment row was unified; dead exports were removed; and the site wordmark settled on leetbit. Behaviour-preserving throughout — the rendered documents are byte-identical except those deliberate changes.
Compass changed from breadcrumbs to a session-path tree. Navigation memory, not site hierarchy.
Real hit-rate numbers behind the path-resolution article's claims, methodology holes included.
Lookup caching landed with short TTLs after the dentry-cache measurements showed how badly naive FUSE lookup performs.
The dentry cache as memoization. Suspiciously clean analogy — currently hunting for what it hides.
First measurement-driven note: forced an iput storm on a small VM and watched the shrinker evict 22k inodes in under a second.
The namei walk, RCU-walk, and the -ECHILD fallback pattern. Became the reference article for the atlas-fs lookup work.
Prose typography, code blocks, footnotes, and references reached the point where two-hour reading sessions stopped being fatiguing.
First long-form article. The lifetime rules turned out to be two mechanisms, not one — the article exists because that distinction kept producing bugs.
Writes would couple the filesystem to a format that is still moving. Read-only keeps the VFS learning focused.
A FUSE filesystem over the knowledge base, started to force a concrete understanding of the VFS contract.
Rejected per-document frontmatter relations after two weeks of drift. Central graph, build-time validation, no exceptions.
First commit of the renderer. The premise: a knowledge base should compile like software, with relations validated at build time.