ZLFS: de-drift the docs; link the repo from the slides

README still described the pre-roadmap state in places: a single-
indirect ~2 MB file cap, whole-file buffering, the copying cleaner and
the concurrency-safe commit as future work, and 'no garbage collection'
under non-goals -- all long since implemented and VM-validated.  Bring
the read/write/GC bullets, the limitations list and the non-goals in
line with the code, and replace the completed 'remaining sequence' with
a pointer to the stacked f1/f2/f3 branches awaiting incremental VM
validation.  docs/zlfs-design.md section 5 similarly still called the
copying cleaner future work; describe the implemented dirty-overlay
compaction instead.

The slides now link the repository (slide 1 eyebrow and a footer line
on the roadmap slide, with the evidence log alongside).

Author: xsub

README.md

@@ -407,13 +407,14 @@ Kernel (`sys/zlfs/`):
   per-zone allocator state, all validated against device geometry with
   every on-disk field bounded before use.
 - Read: a per-mount vnode cache, `lookup`/`readdir`/`read` over on-disk
-  inodes, directories, and both direct and single-indirect data blocks.
+  inodes, directories, and data blocks through the direct, single- and
+  double-indirect pointers.
 - Write: a raw zoned-write primitive (`dk_zone_write_kern`, a direct
   `WRITE(16)` bypassing the buffer cache and the host-managed write gate),
   an append-only log allocator, and `create`/`write`/`fsync`/truncate,
-  `mkdir`/`rmdir`, `unlink`, and `rename`.  Regular files grow past the
-  twelve direct blocks through a single indirect block, and the in-core
-  file buffer grows on demand rather than reserving the maximum up front.
+  `mkdir`/`rmdir`, `unlink`, and `rename`.  Regular files reach ~1 GB
+  through single- and double-indirect trees, and writes land in a sparse
+  per-block dirty overlay rather than a whole-file buffer.
   Directories nest: `..` resolves to the real parent, directory link
   counts track subdirectories, and `rename` reparents `..` (rejecting a
   move of a directory into its own subtree).  A commit flushes all dirty inodes as a
@@ -432,32 +433,32 @@ Kernel (`sys/zlfs/`):
   the durable checkpoint (re-read from disk, not trusted from memory),
   from the in-core inode map, and from in-core vnodes (covering
   unlinked-but-open files), then resets any written zone reachable from
-  none of them.  Only fully dead zones are reclaimed, and any read
-  failure aborts the scan without reclaiming, so a crash at any point
-  still recovers everything the durable checkpoint references.
+  none of them.  Fully dead zones are reset; mixed zones are compacted —
+  the sync path relocates their live blocks through the dirty overlay so
+  the zone goes dead and a later pass resets it.  Any read failure
+  aborts the scan without reclaiming, so a crash at any point still
+  recovers everything the durable checkpoint references.
 
 Current limitations (documented in the code; each is a natural next step):
 
-- Single indirect block only: files are capped at
-  `ZLFS_NDADDR + block_size/8` blocks (about 2 MB at a 4 KB block size);
-  no double or triple indirect blocks yet.
-- Each open file or directory is buffered whole in core, so the maximum
-  file size is also bounded by available memory; there is no block-level
-  buffer-cache integration.
-- The cleaner reclaims only fully dead zones; zones holding a mix of
-  live and superseded blocks are not compacted (a copying cleaner is
-  future work), so space in mixed zones is reclaimed only once
-  everything in them is superseded.
+- Files are capped at ~1 GB (direct + single- + double-indirect);
+  triple-indirect support is implemented and awaiting VM validation on
+  the `zlfs-f1-tripleind` branch.
+- Directories are rewritten whole on commit and capped at
+  `(12 + block_size/8) * block_size` (about 2 MB at a 4 KB block size).
 - Reads go through the buffer cache.  This is coherent with the raw
   zoned writes because the log never overwrites a live LBA; the cache
   is purged whenever a zone reset makes cached blocks stale.
-- The commit path is not yet safe against concurrent vnode operations.
+- Special files, symlinks, hard links, and advisory locks are not
+  implemented (`EOPNOTSUPP`).
 
-Remaining sequence toward a general-purpose filesystem:
-
-1. Double/triple indirect blocks (files past ~2 MB).
-2. A copying cleaner (compact mixed live/dead zones).
-3. Concurrency-safe commit.
+The original bring-up roadmap — indirect blocks, per-block commit, the
+copying cleaner, and the concurrency-safe commit — is complete and
+VM-validated (see `functional_testing.md`).  Work beyond it lives on
+stacked branches awaiting incremental VM validation: `zlfs-f1-tripleind`
+(triple-indirect blocks), `zlfs-f2-powerfail` (power-cut simulation and
+torn-write recovery), and `zlfs-f3-fsck` (`fsck_zlfs(8)`, an offline
+consistency checker).
 
 Regular files no longer buffer their whole contents in memory: writes
 land in a sparse per-block dirty overlay, partially covered blocks are
@@ -467,8 +468,9 @@ file no longer rewrites the file.
 
 ## Non-Goals For The Current Prototype
 
-- Not production-ready: no garbage collection, limited file and directory
-  sizes, and durability caveats above.
+- Not production-ready: experimental on-disk format, limited directory
+  sizes, and validation so far only on QEMU NVMe ZNS (no physical SMR
+  or cache-backed hardware yet).
 - No filesystem-level zoned allocation policy beyond append-at-write-pointer.
 - No promise of on-disk format or ABI stability before review.
 - No attempt to support drive-managed SMR specially; those already appear as normal disks.

docs/zlfs-design.md

@@ -116,8 +116,14 @@ resets any written data zone reached by none:
 3. **Open vnodes**: unlinked-but-open files live in neither map.
 
 Any read failure aborts the scan (reclaim nothing).  The log head is
-never reset while it has room.  Only fully dead zones are reclaimed —
-compacting mixed zones (a copying cleaner) is future work.  After any
+never reset while it has room.  Fully dead zones are reset directly;
+mixed zones go through the **copying cleaner**: when free zones run
+low, the sync path (which holds no vnode locks) picks the least-live
+written zone and pulls every live data block owned by any inode into
+that inode's dirty overlay, marking inodes whose *metadata* sits in
+the victim for a full indirect-tree rewrite (`zn_relocate`) — so the
+next commit relocates everything, the zone goes dead, and a later
+pass resets it.  After any
 reset the device buffer cache is purged (`vinvalbuf`): ZLFS writes
 bypass the cache, and reset+reuse is the only way an LBA's contents
 change, so this single invalidation point keeps cached reads coherent.

docs/zlfs-slides.html

@@ -43,11 +43,12 @@
 .t-next{border:1px solid var(--mut);color:var(--mut)}
 .counter{position:fixed;right:16px;bottom:14px;font:11px ui-monospace,Menlo,monospace;color:var(--mut)}
 code{font:500 .92em ui-monospace,Menlo,monospace;color:var(--acc)}
+a{color:var(--acc);text-underline-offset:3px}
 @media (prefers-reduced-motion:no-preference){html{scroll-behavior:smooth}}
 </style>
 
 <section class="slide" id="s1">
-  <p class="eyebrow">Slide 1/5 · OpenBSD-src fork · xsub/OpenBSD-src</p>
+  <p class="eyebrow">Slide 1/5 · OpenBSD-src fork · <a href="https://github.com/xsub/OpenBSD-src">github.com/xsub/OpenBSD-src</a></p>
   <h1>What is this project for? <em>Native zoned storage for OpenBSD.</em></h1>
   <p class="lead">Modern high-capacity drives — host-managed SMR and NVMe <b>ZNS</b> — refuse random writes: data may only land on each zone's <em>write pointer</em>. OpenBSD had no native support for them at all.</p>
   <div class="cols">
@@ -116,6 +117,7 @@ <h1>Where we are, <em>what remains</em></h1>
     <li><span class="tag t-next">BEYOND</span><span>POSIX gaps — symlinks, hard links, advisory locking (flock/lockf)</span></li>
     <li><span class="tag t-next">BEYOND</span><span>Validation on a physical SMR drive (so far: QEMU NVMe ZNS)</span></li>
   </ul>
+  <p class="legend" style="margin-top:20px">Code, docs &amp; tests: <a href="https://github.com/xsub/OpenBSD-src">github.com/xsub/OpenBSD-src</a> · evidence log: <a href="https://github.com/xsub/OpenBSD-src/blob/main/functional_testing.md">functional_testing.md</a></p>
 </section>
 <div class="counter" id="ctr">1 / 5</div>
 <script>