make: say what the version stamp is, and what it is not The check-version comment claimed the binary "reports itself from the VCS revision Go records". It does not. cmd/specsrht declares `const version = "dev"` and hands it to mcpsrv.Handler, nothing in the tree calls debug.ReadBuildInfo, so every build introduces itself to an MCP client as "dev". The stamp is provenance readable with `go version -m`, which is what this gate is for and all it now claims. Keep the mechanism rather than adopting the siblings' -ldflags: there is no symbol to set. `-X` at a const and `-X` at a symbol that does not exist are both silently ignored on go1.26.5 — exit 0, no diagnostic, the binary still prints "dev" — so a VERSION here would read like a version and do nothing. Record what has to change first, and cov.sr.ht's argument for changing it. Also name the linked-worktree case in the empty-dirt branch, which is how it was found: built from .worktrees/chore/makefile-preflight with a clean tree, the binary came out vcs.modified=true because Go had stamped the parent checkout. Every measurement above was taken in a plain clone for that reason.
cmd: keep the metrics listener on loopback unless asked otherwise core-go defaults the Prometheus bind to ":0" - a random port on every interface - while giving pprof "localhost:0", and this daemon inherited the wrong half of that asymmetry: /metrics is served by core-go's own mux with no credential in front of it, so anyone who could reach the host read the counters of an instance whose service port is deliberately on loopback behind nginx. withMetricsDefault inserts -m localhost:0 immediately after argv[0], so an operator's own -m is parsed after it and still wins - server.New keeps the last -m it sees. The vector is built once in run and handed to both bindAddresses and server.New, so the startup line and the daemon cannot disagree about it.
cmd: take the listen address from the config the operator already edits The daemon's only HTTP listener was set by -b alone, with localhost:5091 compiled in as the fallback, so an instance that needs a different address had to carry it in the unit file or the container entrypoint rather than in the config.ini it edits for every other setting. phoebe-lab's template says so in as many words: "the web listener is set with -b in the entrypoint ... so no bind-address key is needed here", while every containerised sibling beside it binds 0.0.0.0 through its own key. [spec.sr.ht] bind-address is that key. The spelling is not invented here: six siblings already read it under that name -- artifacts, bench, cover, curator, snip and tokens -- and upstream's own builds.sr.ht worker reads [builds.sr.ht::worker] bind-address, so it is the name an operator sharing one config.ini across services already knows. Precedence is flag > config > default, stated in service.DefaultBindAddress and in bindAddresses rather than left to be inferred: server.New falls back to the address it was constructed with only when the vector carries no -b, so handing it cfg.BindAddress is what produces that order. The address is validated while the config is read, next to repos and origin, so a malformed one is refused by a message naming the key instead of by a bare net error once the database is open and the hook socket is held. Port 0 is refused with the malformed ones -- it is legal to bind and useless behind a proxy that connects to a fixed port. bindAddresses reads the vector with core-go's own getopt rather than scanning for "-b", so the attached form and clustered flags are read the way the real parse reads them, and the startup line reports the addresses actually bound rather than the configured one -- which -b had already been making it misreport.
web: say in the embed directive that partials are wanted //go:embed templates/*.html carried _threads.html by accident of a rule rather than by saying so: embed excludes a name starting with '.' or '_' when it walks a directory, not when a pattern matches the file directly. Measured before changing anything — the partial is in the embedded FS today, and TestEveryTemplateOnDiskIsEmbedded is what keeps it there. The "all:" prefix states the rule instead of relying on the reader knowing the distinction, so the line is right under either reading of it and cannot be shortened into the directory form without the prefix being deleted first. The glob stays: "all:templates" is the same intention with no extension to match, and would compile a stray .DS_Store or an editor swap file into the release binary, which pages.Load would filter out in silence. The comment on the neighbouring static directive moves with it, since the two forms now differ on the page: that tree is served to a browser by name, so excluding '.' and '_' is what is wanted there. It also listed two assets where there are three — proposal.html has loaded /static/diff.js since the prose differ grew a client side.
web: serve the asset tree for the two methods it answers r.Mount registers a subtree under every method chi knows, so the routing table listed all ten against a handler that answers two, and a same-origin POST to /static/main.min.<sha>.css came back 200 with the whole 133 KB stylesheet instead of a 405. Nothing was writable and nothing leaked; the description was wrong, and the table is what chi's own 405 and anything walking the router read. chimw.GetHead over the pattern Mount would have appended, so the asset tree is registered the way every other read route on this surface is. Nothing of Mount's is lost: its RoutePath shift and its 404/405 inheritance are for a mounted *Mux, and assets.Handler is a plain handler that reads r.URL.Path and strips the prefix itself. Measured against a live listener, before and after: the hashed asset still answers 200 with public, max-age=31536000, immutable and no Vary, a HEAD still answers the same headers with no body, an If-None-Match: * still answers 304 carrying the asset policy rather than the page's private, no-store, and /static, /static/ and a name that is not there still get this service's own 404 page. POST, PUT, PATCH, DELETE and OPTIONS now get the rendered 405. The 405 carries no Allow header, which is chi's limitation rather than this route's: it hands the list of methods that would have matched only to its built-in responder, through unexported types, so the custom handler chimw.RenderRefusals installs cannot emit one. Every 405 on this surface is missing it, not just this route.
api: tell caches that the write plane answers per credential Every answer the REST write plane gives is computed from the bearer token on the request and from nothing else the request line shows: the 201 names a proposal opened for the agent that presented the credential, the 401 names the challenge, the 403 says this token may not propose. The mux carried no cache policy at all, which made it the last credential-varying surface of the service without one — web/ has it, /mcp has it, and /query got it two commits ago. The policy is sr-ht-ecore's PrivateCache, the same helper the other three use: private, no-store with Vary: Cookie, Authorization. Where it goes is a short question here and was a long one for mountGraphQL, because that Group also holds api-meta.json and had to keep it publicly cacheable. This mux mounts exactly one route and it is the credential-varying PUT, so the policy covers the whole mux and picks up two answers a route-level one would miss: chi's 404 for an address this surface does not serve, and the resolver middleware's 401, written before the router has looked at the path. The honest ranking, which is why this is one line: a PUT is even less cacheable than the POST /query carries the same policy for, so this is defence in depth and not a live leak. The test drives httptest.NewServer and a real client rather than a recorder, which hands back the live header map and so cannot observe when a header was set
doc: check the href a review link publishes, not the destination it came from Mutation found the dangerous-URL refusal had grown two guards for one property: with an archive supplied, the check on the way out caught what the check on the way in was for, so removing the incoming one turned nothing red. There is one check now, on the href, which is the thing that reaches the anchor — and removing it fails four tests in both configurations. The relative-link test went the same way: a sibling destination resolves through doc.Archive's space-wide stem fallback as well, so it passed with the linking document's directory thrown away. It uses a path-qualified destination now, which has no fallback.
web: resolve a proposal's links against the revision under review An internal link inside a proposed document used to reach the review page as the destination as typed, so following one landed nowhere. The page now reads the proposal through ProposalReview and resolves each block against the archive of the revision that block's text is from: the branch tip for the text under review, the base for a block the proposal deletes. Every href is pinned to the revision it was resolved at. A document's site path names the document and not the revision, so an unpinned link out of a proposed block would land on the approved head — the text the proposal is changing. Two things the resolver could not say before and says now: a link to a document that does not exist is marked, and so is a citation of a section the proposal has renamed, which is the error a proposal introduces and review is the moment to catch.
service: read a proposal's two revisions as archives ProposalDiff hands over bytes and a path, so nothing above it can turn a wikilink in a proposed document into a URL. ProposalReview returns the same documents plus the whole document set of each side as an archive, built from the tree walks the diff already does. Two archives and not one because the review page renders two revisions: a block the proposal deletes belongs to the base, and resolving its links against the branch tip would answer for it with the very text the proposal is changing. Each is built once per proposal rather than once per document; LinkPass is a parse of every document in the revision. ProposalDiff is left as it was, so the comment POST path and the MCP surface pay for neither.
doc: let the review renderer resolve against a supplied archive Inline now takes the linking document's directory and the document set to resolve against, and wraps whatever it is handed in reviewResolver rather than holding one. The wrapping is the security boundary: doc.Archive applies no dangerous-destination check, correctly, because the read plane's documents are approved and render with unsafe on. web/ still passes nothing, so the review page is unchanged by this commit.
web: describe the selection the composer actually gets The composer's note explained that the script prefixes a selected line range onto it and removes that element on Escape. Nothing prefixes anything any more: the gutter drag is gone and a selection becomes a quotation in the textarea. The comment now says that, since it is the one place a reader looks to learn what the composer promises before they type.
scss: stop calling the margin a flex row, which it is not
scss: mark the block being quoted, not the empty half of its cell The tint that says "this is the block your comment will attach to" was painted on the whole text cell. On a full-bleed review that cell runs about 500 pixels past the end of the measure and another 200 past the margin the conversation sits in, so the mark was mostly a wash of colour over nothing — which reads as a rendering defect rather than as a mark. It is on the block's own text now, plus the two number cells, which is what "this block" means. The conversation beside it is left alone: a thread is not part of what is being quoted, and this replaces the rule that had to paint the page's background back over it.
web: pin the placement a stranded conversation falls back to writeGroup pairs a notes row with the row before it only when the two belong to the same block, because the other reading — "put this conversation on whatever row is above it" — files a critique under a paragraph nobody wrote it about. The pairing holds by construction today, which is exactly why the branch that does not was untested: nothing on the page reaches it. It is reached here directly, with a group holding the notes row alone, and what it has to prove is that the comment is still on the page.
scss: give a phone its comment back from the gutter The conversation lives in the block's text cell now, so on a phone it starts where the text starts — 88px in, behind two number tracks and a sign column. Measured on a 375px viewport: 257 pixels of comment on a 375 pixel screen. The full-width notes row this replaced had a rule for exactly that, and it says why: a review comment is the one thing on the page that has to stay readable at any width. So the rule moves to where the conversation moved. Below Bootstrap's sm breakpoint the box is pulled back over the rail, which is empty at that height — the gutter's number sits at the top of the row and the conversation hangs off the bottom of it. 345 pixels of the 375 afterwards, and the page still does not scroll sideways.
web: quote the text you selected instead of dragging line numbers Dragging the gutter was how a reviewer said what they were talking about, and it stopped meaning anything when prose started flowing: a paragraph is one row with one range in the gutter, so a drag could only ever say "this paragraph", which is what the composer already says. Selecting the words says which words. They land in the composer as a `> ` quotation — something the reviewer can edit, delete or keep, and which survives into the posted comment, where it says what the critique is about to whoever reads the thread later. The anchor does not move. A comment is still stored against its block's content hash, heading path and index, which is why the selection is still CLAMPED to one block: a quote reaching into the next paragraph would have to be filed under a paragraph it is not about. The clamp does not argue with the mouse — it takes the part of the selection lying inside the block the selection began in, and the quote in the composer is the report of what was taken. Measured: a 701-character selection across two blocks quotes exactly the first block's 163 and opens that block's composer. A quote is not a draft. `hasDraft` ignores a body that is exactly the quotation this file inserted, and without that every stray selection would leave a composer that nothing could ever close — "one composer at a time" would decay into a box beside every block, which is the state the whole hiding rule exists to prevent. Type anything beside the quote, or edit the quote itself, and it is a draft like any other: measured, the second selection closes the quote-only composer and clears its body, and leaves the one holding words open with the words in it. The address bar carries `#b-<digest>` — the block, not a range. That id is honest across revisions by construction, since the digest covers the block's content hash, so a link into a block that has since been rewritten resolves to nothing rather than to its neighbour. What is lost is a shareable sub-block selection; the quote in the comment body carries those two sentences to the reader of the thread instead of to the follower of a link. A character range under the same id would in fact be honest for the same reason, and is written down at the top of diff.js as the option not taken rather than left to be rediscovered. The gutter cells and the sign are stripped out of a quote. That matters in one narrow case, measured with the filter removed: a drag beginning above the diff is read from its end, so the clamp opens the range before the block's row and the quote comes out as `> 15–2015–20Спека даёт боту…`. Selecting somebody's posted comment does nothing at all, deliberately: quoting a critique into the box that comments on the paragraph would be an answer to the wrong thing.
web: move a block's conversation into the review page's margin A thread was a full-width row under its block, so every critique, every reply and every compose form pushed the two halves of the document apart by its own height — on a page whose subject is how those paragraphs read together. It now sits beside the block instead, out in the empty right-hand third of a full-bleed review. THE MARGIN IS A CELL AND NOT A COLUMN, which the table forces. The diff has to be a table (a paragraph is six visual lines and its numbers have to stay on the first of them), and a table has no column a row can opt into: a rowspanning fifth column would have to survive a fold cutting a block in half, which it cannot, and a fifth column on every row would take its width from the text on a phone. What a table does have is a cell that can hold two things. So the conversation is written into the block's own text cell and the stylesheet sets the two side by side. What that buys is that a row is as tall as its tallest cell: a thread beside a paragraph now costs the document nothing. What it costs is a conversation TALLER than its block — a heading with a long critique on it — which still pushes what follows down, by the difference rather than by the whole. Two inline-blocks and not a flex row. `display: flex` on a td is the obvious spelling and it is wrong: the cell stops being a table cell and stops driving its row's height, and measured in Chrome 151 every commented block painted over the three rows below it. A code fence keeps its conversation underneath, because only its last line would make room and one short line in a listing is a lie about where the scrolling starts. Below 1200px the two stack and the thread is back under its block, which is also what it does with JavaScript off. The row model is untouched: writeGroup folds the notes row into the row before it, the two are adjacent and share a fold by construction, and a notes row that ever arrived without its block in front of it still renders as the full-width row it used to be. The deleted-block strike moved from the cell to the block's own div. It had to: text-decoration propagates and no descendant can switch it off, so a strike started on the cell runs through the comments written about the paragraph. Measured both ways — with the strike back on the cell Chrome 151 leaves the wide layout alone (an inline-block is an atomic inline and exempt) and strikes the comment's author, timestamp, anchor and body in the stacked one. The rule that keeps typed text on the page was ported first. A composer holding a draft inside a fold that is then shut now keeps the block's whole row drawn rather than a notes row of its own; measured with the rule removed, the textarea collapses to 0px with the words still in it.
web: say which of the two guards keeps a comment off the map Mutation-testing the tick sizes turned up a property with two guards and a test that cannot tell them apart. mapTicks skips a notes row when it adds up a block's lines — and deleting that skip changes nothing, because a notes row carries no gutter number on either side and so already reports zero source lines through diffRow.sourceLines. Both the check and its test are kept: what would break the property is a notes row that grows a number, and then the skip is what holds and the test is what says so. What is not kept is a comment implying the skip is doing the work.
web: give the review page a map of what changed The page could say how many blocks changed and could show each of them, and still refused the question a reviewer arrives with: is this a new document, a two-word fix, or a rewritten chapter. The map answers it in shape. One tick per block down the right edge, coloured by what happened to that block, so a proposal that adds a document is a solid bar, a one-word edit is a single coloured tick in a grey column, and a reworked section is a band. Context blocks are ticks too. The coloured ones are only readable against how much document they sit in — a map of the changes alone draws the same picture for both ends of that scale. The ladder in tickWeight is deliberately not linear: a document with one forty-line fence in it would otherwise be a bar of code with every paragraph beside it too thin to see. A changed block is a link to itself and an unchanged one is a mark on the ground, which bounds the links by the size of the change rather than by the size of the document. A tick that points into a collapsed run opens it: the new `:has(:target)` rule is what keeps that working with JavaScript off, since a browser cannot scroll to a row it is not drawing. Nothing is derived twice. rowBlock now carries the block's own change kind, because a move-in opens with a marker row and a modified fence opens with one too, so reading the kind off a block's first row would colour a rewritten fence as a move. mapTicks reduces the same row stream the table is built from, so the map cannot claim a block the table does not draw. The map is aria-hidden and out of the tab order: every tick points at a block that is already on this page in this order, and a new document would otherwise put a hundred empty links in front of the keyboard. The stats line above says the same thing in words. diff.js adds the one thing markup cannot, the scroll indicator, by observing the block rows themselves — with the sticky heading band cut out of the root. Without that, a `tr.ph-head` pins at top: 0 and never leaves the viewport: measured on a 5000px render of one RFC, scrolled to the bottom, nine of the fifty-one ticks were still lit and every one of them was a heading long since passed. Neither getBoundingClientRect nor offsetTop can tell a stuck row from one that is genuinely at the top.
hooks, mcpsrv: stop two comments and one message from naming stale facts The push refusal spelled "spec:propose" as a literal — the one line an agent reads before going to mint a token, so a rename that missed it would send the agent after a grant this service no longer honours. It now names authn.ActionPropose, which the grant vocabulary test holds the docs to. mcpsrv/cache.go said mcphttp was "not resolvable: the ecore commit that adds mcphttp is not published". It is in the ecore this module pins today. What still holds the swap is that mcphttp's vary is a package constant, so the comment says that instead.