Yesterday's announcement ended with a list of things not claimed. The first item on it was the write-back cycle: committing a memory-resident database back to disk, designed and not built. It is built. The more instructive part of the day is that the test guarding it was green before it deserved to be.
What landed
The return leg. A workspace that lives in a memo, or in memory, can now be written back onto a real disk root. The enumeration comes from the posture -- the record of what the workspace is -- never from whichever tables happen to be attached at the moment, because the first cut asked the session and silently wrote fifteen of twenty-seven files while reporting cheerful success. Each landed file is re-read and byte-compared as it is written. A shortfall aborts having written nothing, and an abort leaves the filesystem untouched including empty directories, which had to be fixed after the first version created its target folders before deciding it could not fill them.
The whole cycle, proven by value. Thirteen tables were hydrated out of a memo field into memory with no disk reads, mutated in memory only, saved back into a new container, and landed on disk -- with the memory-only edit read back at the far end. The mutation is the point: it is a value that exists nowhere on disk, so a passing result cannot have come from the original files.
Replacement, and the undo beside it. Writing into an occupied directory refuses by default and lists what it would replace. With confirmation it replaces, and every replaced file is kept alongside the new one. The backup is not asserted to work -- it is opened as a table and read: two hundred records, holding exactly the content that was overwritten.
A load that refuses rather than half-restores. The other end of the same lane had the opposite manners. Given a posture whose tables had moved, the loader closed every open area, failed every re-open, and finished on a sentence containing the word restored. Every individual failure printed, so the transcript was honest line by line and misleading as a whole -- and the last line is what a script reads. It now resolves every declared table before it closes anything, so a load that cannot complete leaves the workspace you already had standing. Restoring only what exists is still available, but you have to ask for it by name.
The green that was not true
The regression guarding write-back reported six passing markers off a directory it had never written to.
Two things that looked identical were not: the verb resolved a relative target against the process working directory, while the command that pointed the reader at that target resolved the same text against the data root. They named the same place only because the standard test runner happens to start in the data root. Run from anywhere else, the test wrote to one directory and read from another -- and the directory it read had been left populated by an earlier, unrelated run. It reported success by consuming someone else's output.
The spec had a guard against exactly this. Its own header warned about stale directories producing false greens, and it ran a pre-clean to prevent them. The pre-clean reported success. It cleaned the directory the reader never touched, because the guard resolved paths the same wrong way as the thing it guarded.
That is the part worth keeping. A guard that shares its subject's defect does not fail loudly -- it agrees. Two components agreeing is not evidence that either is right.
It surfaced only when the stale directory was deleted as housekeeping and the same test, on the same binary, went red. Nothing about the code under test had changed; the accidental input had gone away. The four surfaces now resolve relative paths by one rule, the false transcript is retained and marked as evidence for its own retraction rather than deleted, and the record that cited it has been corrected.
What is not claimed
The backup kept beside a replaced file is one generation deep and kept indefinitely. A second confirmed write-back discards the first backup without saying so. That was measured rather than assumed, and it is a retention policy nobody has chosen yet.
Index environments that must map a real operating-system file are still not carried; the destination rebuilds them. Two workspaces still cannot be resident at once, for the same naming reason as before. Hydration still writes its files one at a time, so a failure part-way leaves a partial set in memory -- acceptable only because that set is disposable, and the reason write-back gathers everything before writing anything.
One open hazard is under review rather than fixed: the directory scanner admits the backup files as tables. A workspace opened over a directory that has been written back twice reports twenty-six tables where the posture declares thirteen, and each backup attaches the live table's index -- an index built over different records. The two features are individually reasonable and jointly wrong, which is the same shape as the path defect above.
And one tool is broken by this work: the metadata collector no longer links, because a function added for the tests took a dependency the collector's library does not carry. It defaults to off, so every build and every gate passed while it was unbuildable. One catalog is stale by exactly that function until it is repaired.
Details: Workspaces · Proven capabilities · Schemas: a table of databases