Nothing couples emission of TAG_SYMBOL_BASENAME to the fact_minor_min declaration that is supposed to bracket it #280

Closed
opened 2026-09-16 09:24:04 +02:00 by buildagent · 1 comment
Member

Split out of #277, whose disclosure fix landed in d5547cc. This half was NOT fixed there — it has a different fix location and deserves its own decision.

The claim

crates/abi/src/frame.rs argues that TAG_SYMBOL_BASENAME (0x8004) is the one optional tag that cannot degrade safely, because the field it annotates — a symbol's name — has no absent encoding. An older host skipping it does not thin the answer; it stores the guest's PLACEHOLDER as if it were a name. The doc's word is that it MISLEADS, which is worse than degrading and worse than refusing.

_prdoc/guides/80-threat-model.md §4.2c states the mitigation as a property of the system:

A package emitting it declares fact_minor_min = 2, and a host below that refuses the package at install with manifest.abi_unsupported rather than storing the placeholder.

It is a property of a COOPERATING package

Manifest::validate refuses fact_minor_min > ABI_MINOR — a package cannot ask for a host newer than the one it meets. Nothing anywhere ties EMISSION of the tag to the DECLARATION.

A third-party package that calls Encoder::symbol_basename while declaring [abi] fact_minor_min = 0 installs happily on a host at ABI_MINOR = 1, whose decoder takes the t if t >= TAG_OPTIONAL_BASE => continue arm and stores the placeholder verbatim as a symbol name. That is precisely the case the bracket is documented as preventing.

Nothing in the tree grades this. Dashboard.expected says so itself: "The expectation format has no field for the annotation itself (ExpectSymbol is deny_unknown_fields), so this table cannot assert that the tag was attached." The tag's presence is graded only by the daemon e2e; the ABI bracket is graded nowhere.

Scope

No shipped package is affected. de.h-dv.svelte declares fact_minor_min = 2 correctly, and it is the only emitter. This is a trap for a well-meaning third-party author, and — more importantly — a sentence in the threat model that reads as a system property when it is a convention.

The related finding from the same review: crates/guest/src/frame.rs's ABI_MINOR is 0, so a guest stamps minor 0 into every frame it writes while those frames may carry a minor-2 record. frame.rs's "an older parent meeting a newer worker is the refused case" therefore never fires for this tag, because the worker never declares itself newer. wire_parity.rs records the non-move as deliberate (moving it moves PINNED_REV), and 09305a3 shows that pin does get moved when needed — so "it is expensive" is not the whole answer.

Where the fix belongs

plugin check's conformance leg. It is the one place that holds BOTH the decoded ValidatedFacts — hence Symbol::name_is_basename — and the Manifest. A package whose frame carries the tag while its manifest declares fact_minor_min < 2 should fail conformance, which means it can never reach plugin enable (which requires the verdict).

Two alternatives worth weighing rather than assuming:

  1. Refuse at validate. Cleanest in principle — the decoder knows the tag arrived. It does not know the manifest: abi::validate takes (body, src, limits, grants) and has no manifest by design, for the same reason it refuses a received line. Handing it one would be the wrong shape.
  2. Refuse at plugin pack. Catches it at authoring time, which is friendlier. But packing does not decode a frame, so it would have to run the extractor — which is what plugin check is for.

(1) is the tempting one and (2) is the friendly one; the conformance leg is probably right because it already executes the package and already gates enable.

Also from #277, and NOT a defect — a decision someone should make

Should GrantSpec::Requested grant derived_names at all by default? plugin add with no --grant resolves it to whatever the manifest asked for. After d5547cc the confirmation screen NAMES it, so the reported hole is closed. Every other high-authority answer in this tree is opt-in, so "everything the manifest asked for, including this" is now defensible but no longer obviously right. Changing it is a behaviour change for anyone scripting plugin add, which is why it was not bundled into the disclosure fix.

Split out of #277, whose disclosure fix landed in `d5547cc`. This half was NOT fixed there — it has a different fix location and deserves its own decision. ## The claim `crates/abi/src/frame.rs` argues that `TAG_SYMBOL_BASENAME` (0x8004) is the one optional tag that cannot degrade safely, because the field it annotates — a symbol's `name` — has no absent encoding. An older host skipping it does not thin the answer; it stores the guest's PLACEHOLDER as if it were a name. The doc's word is that it **MISLEADS**, which is worse than degrading and worse than refusing. `_prdoc/guides/80-threat-model.md` §4.2c states the mitigation as a property of the system: > A package emitting it declares `fact_minor_min = 2`, and a host below that refuses the package at install with `manifest.abi_unsupported` rather than storing the placeholder. ## It is a property of a COOPERATING package `Manifest::validate` refuses `fact_minor_min > ABI_MINOR` — a package cannot ask for a host newer than the one it meets. **Nothing anywhere ties EMISSION of the tag to the DECLARATION.** A third-party package that calls `Encoder::symbol_basename` while declaring `[abi] fact_minor_min = 0` installs happily on a host at `ABI_MINOR = 1`, whose decoder takes the `t if t >= TAG_OPTIONAL_BASE => continue` arm and stores the placeholder verbatim as a symbol name. That is precisely the case the bracket is documented as preventing. Nothing in the tree grades this. `Dashboard.expected` says so itself: *"The expectation format has no field for the annotation itself (`ExpectSymbol` is `deny_unknown_fields`), so this table cannot assert that the tag was attached."* The tag's presence is graded only by the daemon e2e; the ABI bracket is graded nowhere. ## Scope **No shipped package is affected.** `de.h-dv.svelte` declares `fact_minor_min = 2` correctly, and it is the only emitter. This is a trap for a well-meaning third-party author, and — more importantly — a sentence in the threat model that reads as a system property when it is a convention. The related finding from the same review: `crates/guest/src/frame.rs`'s `ABI_MINOR` is **0**, so a guest stamps minor 0 into every frame it writes while those frames may carry a minor-2 record. `frame.rs`'s "an older parent meeting a newer worker is the refused case" therefore never fires for this tag, because the worker never declares itself newer. `wire_parity.rs` records the non-move as deliberate (moving it moves `PINNED_REV`), and `09305a3` shows that pin does get moved when needed — so "it is expensive" is not the whole answer. ## Where the fix belongs **`plugin check`'s conformance leg.** It is the one place that holds BOTH the decoded `ValidatedFacts` — hence `Symbol::name_is_basename` — and the `Manifest`. A package whose frame carries the tag while its manifest declares `fact_minor_min < 2` should fail conformance, which means it can never reach `plugin enable` (which requires the verdict). Two alternatives worth weighing rather than assuming: 1. **Refuse at `validate`.** Cleanest in principle — the decoder knows the tag arrived. It does not know the manifest: `abi::validate` takes `(body, src, limits, grants)` and has no manifest by design, for the same reason it refuses a received line. Handing it one would be the wrong shape. 2. **Refuse at `plugin pack`.** Catches it at authoring time, which is friendlier. But packing does not decode a frame, so it would have to run the extractor — which is what `plugin check` is for. (1) is the tempting one and (2) is the friendly one; the conformance leg is probably right because it already executes the package and already gates `enable`. ## Also from #277, and NOT a defect — a decision someone should make Should `GrantSpec::Requested` grant `derived_names` at all by default? `plugin add` with no `--grant` resolves it to whatever the manifest asked for. After `d5547cc` the confirmation screen NAMES it, so the reported hole is closed. Every other high-authority answer in this tree is opt-in, so "everything the manifest asked for, including this" is now defensible but no longer obviously right. Changing it is a behaviour change for anyone scripting `plugin add`, which is why it was not bundled into the disclosure fix.
Author
Member

Fixed in PR #491 and released in v0.41.36.

plugin check now refuses observed basename annotations when fact_minor_min < 2, with conformance.fact_abi_undeclared and a witness. Verdict schema 3 invalidates lower-floor legacy passes until rechecked; verified packages already declaring minor >= 2 preserve their legacy verdict. ABI decoding, SDK pins and grant defaults are unchanged. The threat model states the fixture-based scope explicitly; this cannot retrofit older hosts or prove untested inputs.

Regression tests failed before the fix and pass now. Real signed-package public CLI tests passed again against the installed release: minor 0/1 basename packages are refused, minor 2 passes and indexes the filename-derived name, and the unchanged minor-0 Timeline package passes after recheck and re-enables. Linux (5,358 main + 53 timing), Windows (5,132 main + 43 timing), package/pilot qualification and all ten release jobs are green. Published archives, signatures, channel preservation and installed bytes were independently verified. The installed release also indexed its own exact published tree using all eleven official packages and passed nine tool checks over both transports.

Fixed in PR #491 and released in [v0.41.36](https://git.h-dv.de/h-dv/code-index/releases/tag/v0.41.36). `plugin check` now refuses observed basename annotations when `fact_minor_min < 2`, with `conformance.fact_abi_undeclared` and a witness. Verdict schema 3 invalidates lower-floor legacy passes until rechecked; verified packages already declaring minor >= 2 preserve their legacy verdict. ABI decoding, SDK pins and grant defaults are unchanged. The threat model states the fixture-based scope explicitly; this cannot retrofit older hosts or prove untested inputs. Regression tests failed before the fix and pass now. Real signed-package public CLI tests passed again against the installed release: minor 0/1 basename packages are refused, minor 2 passes and indexes the filename-derived name, and the unchanged minor-0 Timeline package passes after recheck and re-enables. Linux (5,358 main + 53 timing), Windows (5,132 main + 43 timing), package/pilot qualification and all ten release jobs are green. Published archives, signatures, channel preservation and installed bytes were independently verified. The installed release also indexed its own exact published tree using all eleven official packages and passed nine tool checks over both transports.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
h-dv/code-index#280
No description provided.