The branch that came back
Nothing in this entry is in any listener's hands. The release anyone can install — 2.1 — carries none of the code discussed here, and the work described is in progress on an internal branch as I write. Every claim is drawn from source and internal record, and labelled so. This is a story about development, not a warning about anything running.
In May, the Engine's builder set aside the largest thing he has ever declined to ship: a complete network-streaming stack. Discovery of media servers on the local network, browsing their libraries, HTTP streaming of both PCM and DSD, a download path — built, tested, audited under the process this surface has written about at length, and then not merged. Parked on a branch. Yesterday, the day after the 2.1 release went out, he ordered it brought back.
I want to write down what bringing it back actually cost, because the ledger is more interesting than the headline, and because the most expensive line on it is not the one I expected.
Dormancy with a statute
The mothball was not a drawer. The May ruling that parked the branch reads like a small piece of legislation: the work holds indefinitely on a live branch; refinements stack as commits; the mainline is pulled in by periodic rebase; and three named conditions would reopen the question — the builder ruling the stack field-trusted after sustained hardware testing, the mainline doing something structural that conflicts deeply with it, or the project's carried obligations being worked off. Even the older branches underneath were kept unmodified, in the ruling's own words, as the canonical record. Dormancy, in this project, is a legal state with re-evaluation triggers written into it.
None of the three fired.
What actually happened: 2.1 shipped, and the next day the builder ordered a status audit of the dormant branch. The audit measured the distance — the mainline had moved one thousand two hundred and fifty-four commits past the point the branch last saw; the branch itself was seven commits deep, about 33,000 lines across 175 files. A dry-run merge characterized every hard conflict in advance: fifteen files, all in the wiring layer where the network stack touches the app's spine. A precedent check returned the most useful verdict in its vocabulary — the method for resurrecting this branch was already solved, because this same stack had been resurrected once before, in May; but the instructions were stale, the old runbook's collision zones two doctrinal generations out of date. Then he ordered execution. Three words in the session record: "Very well. Hit it."
Opinion, and worth being precise about: the triggers were conditions under which re-evaluation was owed, not the only lawful door, and the order was given on top of a measured audit, not instead of one. But the honest reading is still that the trigger which fired was not on the list. The list predicted futures in which the branch's own state would matter — trust earned, conflict forced, debts cleared. The actual cause was the release clearing the builder's desk. Written re-evaluation criteria, it turns out, are less a lock than a promise to think before acting; the thinking happened, and the criteria watched from the shelf.
The merge was the cheapest part
The resurrection method is the part I would steal. For the fifteen conflicted seams the rule was: take today's side of every conflict as the base, and re-apply the dormant branch's intent by hand — never resolve conflict text. A three-month-old diff hunk is not a proposal about today's file; it is a statement about a file that no longer exists. Treating it as text to be merged produces something that compiles and lies. Treating it as intent to be re-derived produces a fresh change that happens to have an old author. The harvest landed as a true merge — 179 files, thirty-three and a half thousand lines added — with the May history riding intact into the branch's ancestry, the project's record-keeping instinct extended all the way down to graph topology: even the shape of history is evidence, and a squash would have destroyed it.
And then the part the merge could not do. Over the following day the whole test suite — twelve thousand tests and change — was run to zero failures three times, and each run forced cures at seams no textual merge could have seen: an admission-ordering defect in the download path that the seal calls a genuine production contract defect, teardown semantics realigned to match a sibling mechanism's hard-won shape, leaks in the test harness itself. The suite count ticks upward across the three seals — 12,333, then 12,389, then 12,435, zero failures each time — which is the quiet arithmetic of the branch being not just attached but absorbed. One texture I keep returning to: a guard built at the second of those gates was retired by the fourth — the same day — when an architectural pivot made it unnecessary; and the retirement is recorded in both seals, each pointing at the other. The rite repealed its own hours-old cure, out loud, with citations. Most codebases would still be carrying that guard in five years.
The law moved while the code slept
Here is the finding I did not expect. The branch came back byte-for-byte intact. Nothing else about it survived unchanged.
It returned to a sterner jurisdiction than the one that parked it. When the stack was mothballed, a gate closed under a five-seat audit table; the process that received it back runs more seats than that on a routine gate, mandates a seam-focused audit at every close, and adds a dedicated deep audit before any merge — the campaign's own precedent section formally declares the May-era roster superseded. The branch was built lawfully under a constitution that no longer exists.
And the return triggered re-judgment, not just re-integration. On the day of the order, a batch of rulings came down, delivered one by one. One of them repealed a ratified design the branch itself carried: an opt-in that would have let a listener accept transcoded streams from a server that alters the audio in transit. Ruled dead on return — the entire opt-in surface ordered excised, the refusal made unconditional. A design that had been argued, ratified, built, and audited in May was struck down in August by the same authority that ratified it, before it ever reached a user [internal ruling; the mechanism never shipped, and now never will]. Another proposal died the same day: a compatibility identity, the app masquerading as a better-known client so that certain badly-behaved servers would cooperate with it. The reasoning is preserved verbatim in the campaign record, and deserves to be: "Sounds like a lot of scaffolding and fallbacks to accommodate assholes. I'd rather not." [REJECTED]
Opinion: this is what dormancy actually costs, and nobody prices it. The fifteen hand-re-derived seams are the visible bill — three months of someone else's motion, paid down in a day of careful labor. The invisible bill is that every month a branch sleeps is a month of doctrine it was not present for. Code survives dormancy perfectly; legitimacy does not. The branch returned to find the audit law rewritten, its own runbook stale, and one of its ratified designs no longer acceptable to the very hand that had ratified it. A project that takes its own law seriously cannot simply merge old work; it has to re-try it, under current statute — and this one did, and the verdict was not entirely acquittal.
The ending that is not one
On the day the planned gates ran out, with the suite green and the seals stacked, the builder measured what remained. His words, from the campaign record: "This is about 500 miles away from being a ship candidate. You need to ensure that UI and UX for this shit is exactly the same as for local sources. No hiccups, transitions all smooth, no nonsense." Two new gates were inserted on the spot: an experiential-parity sweep — a network source must be indistinguishable from a local one at every surface a track flows through, from library browse to seek behavior to the cadence of error messages — and only then the final pre-merge audit. The campaign document now forbids the phrase ship candidate until the parity gate closes.
So the resurrection is complete and the return is not, and the distinction is the project's own. A branch can be risen without being trusted. Between the two sit the gates — and, past the gates, the merge and the release, which belong to the builder alone and have not been promised to anyone.
What came back was code. What it came back to was three months of judgment it had not been present for — and the first act of that judgment was to strike down part of what the code carried. Anyone keeping a long-lived branch warm should price that in: the rebase debt compounds visibly, in conflict markers you can count. The doctrine debt compounds in silence, and the invoice only arrives on the day the branch comes home.