Give the audit's open findings a home and a trigger (V-674)

Nineteen of the twenty findings were open, and they lived in an untracked
audit.md at the repo root that no next session would have read. The one that
is closed, the unauthenticated mavgpud proxy, went out as V-673.

The report is now a frozen measurement under docs/evals/, dated and never
edited again — including when a finding it names gets fixed. The live state
moved to docs/caveats/, one entry per limit, each carrying its Vikunja id and
the condition that makes it worth fixing. A caveat with no revisit trigger is
a complaint, so every entry has one. Closing a limit deletes its entry rather
than editing the measurement that found it.

Two directory indexes come with it. docs/CLAUDE.md states the tier rule the
repo already followed by convention: living docs corrected in place, evals
frozen by date, caveats deleted when fixed. docs/caveats/CLAUDE.md indexes the
nineteen by claim and severity, because an index of filenames adds nothing a
directory listing does not.

Tasks V-675 through V-693 carry the plans. The doc line and the tracker now
join in both directions.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ESv8hqNPseYt1CnotZpqDz
This commit is contained in:
2026-08-11 10:41:54 +04:00
parent a1d018dc47
commit c0f4074a5d
12 changed files with 561 additions and 0 deletions
+47
View File
@@ -0,0 +1,47 @@
# docs/caveats/
One entry per known limit: something broken, deferred or unsafe that a session
will otherwise walk into. An entry names what fails, who it costs, and the
condition that makes it worth fixing.
Two things do not belong here. The evidence is a dated file under `docs/evals/`.
The reasoning behind a subsystem is its living doc directly under `docs/`. A
caveat is the pointer between them plus the trigger.
## Rules for this directory
* One file per area, one `##` section per limit, each carrying its task id.
* **A caveat with no revisit trigger is a complaint.** Give it one or delete it.
* Closing a limit deletes its entry. It does not edit it to say "fixed", and it
never edits the frozen measurement it came from. The durable record of a fix
is the commit and the subsystem's living doc.
* An entry whose task is closed but whose limit is still live is the failure
mode to watch for. The id joins the two directions, so check both.
## Index
Every entry below came from the 2026-08-10 deep audit
(`docs/evals/2026-08-10-repo-audit.md`). One of the twenty findings, the
unauthenticated mavgpud proxy, was fixed as V-673 and has no entry.
| limit | severity |
| --- | --- |
| [Go 1.25.5 and x/text 0.14.0 carry 20 reachable advisories](dependencies.md#toolchain) | high |
| [Anyone past the proxy can enroll a passkey](security.md#enrollment) | high |
| [Passkey credentials are rewritten in place](security.md#credentials) | medium |
| [An empty STT transcript reads as a successful one](external-inputs.md#stt) | medium |
| [Open-Meteo's empty body becomes 0°C](external-inputs.md#weather) | medium |
| [Dialogue persistence errors are swallowed](storage.md#dialogue) | medium |
| [The reminder transition is a lost update](storage.md#reminders) | medium |
| [A recall miss scans two whole tables](storage.md#recall) | medium |
| [Closing a TCP listener can strand Accept](transport.md#accept) | medium |
| [PTT reads an unbounded body](transport.md#ptt) | medium |
| [mavweb errors cannot be traced](transport.md#errors) | medium |
| [Fact enrichment is a 20-call serial waterfall](workers.md#enrichment) | medium |
| [A suppressed nudge is phrased anyway](workers.md#nudges) | medium |
| [heads_path may equal model_path](invariants.md#heads) | medium |
| [baselineGrammars is mirrored by hand](invariants.md#grammars) | medium |
| [Committed absolute paths pin the build to this box](config.md#paths) | medium |
| [The env example omits deployed variables](config.md#secrets) | medium |
| [Domain packages depend on store and IPC types](layering.md#dtos) | low |
| [Eleven symbols are unreachable](layering.md#deadcode) | low |
+21
View File
@@ -0,0 +1,21 @@
# Configuration and environment
## Committed absolute paths pin the build to this box [#690] {#paths}
Costs: `go.mod` replaces Hexis with `/home/kami/apps/hexis`, `start-maven.sh`
hardcodes the checkout and the data directory, and `deploy/mavgpud.json` holds
workstation model and Python paths. Vendoring hides the `go.mod` problem for an
ordinary build. `-mod=mod`, `go mod tidy` and a fresh checkout all fail.
Revisit when: anyone clones this repo elsewhere, or a `tidy` is needed.
Workaround: build only from this checkout, with the vendor directory.
## The env example omits deployed variables [#691] {#secrets}
Costs: a fresh deploy can lose remote speech-to-text or ambient authentication
and run on fallback behaviour with an apparently valid config. Three variables
are referenced and undocumented: `MAVEN_STT_TOKEN`, `MAVEN_AMBIENT_TOKEN` and
`CW2_TOKEN`. V-673 added `MAVEN_GPU_TOKEN` to the example.
It is not silent. The loader logs which variables were unset and says whatever
they configure is off. What is missing is a startup failure.
Revisit when: the box is redeployed from scratch, or a new secret is added.
Workaround: read that log line at startup.
+17
View File
@@ -0,0 +1,17 @@
# Dependencies
## Go 1.25.5 and x/text 0.14.0 carry 20 reachable advisories [#682] {#toolchain}
Costs: `govulncheck` found 20 reachable advisories, one in `x/text` and 19 in
the standard library. They include template XSS, parser denial of service and
TLS issues. Reachable traces run through the ONNX embedder's normalization,
mavweb's HTML rendering, email header decoding and the mavgpud proxy. The
vendored toolchain was built 2025-11-26.
Revisit when: now. This is the highest-severity open entry and the fix is
mechanical, so it ages badly for no reason.
Workaround: none.
None of `staticcheck`, `govulncheck` or `deadcode` is installed on this box or
wired into a make target. `make audit` is a git-grep inventory over loc, todo,
stubs, docs, tests and gaps. **Do not read it as a static-analysis gate.** That
gate is part of this entry.
+20
View File
@@ -0,0 +1,20 @@
# External inputs
What arrives from a service Maven does not run, and what happens when it
arrives malformed. The shared shape: a JSON decode into value fields cannot
tell "absent" from "zero", so a degraded response becomes a confident answer.
## An empty STT transcript reads as a successful one [#675] {#stt}
Costs: one dropped voice turn per malformed 200 from workpc. The mavsttd floor
is never asked, because `stt.Pair` falls back on a non-nil error alone.
Revisit when: CW2 returns a 200 with no text. Sooner if a proxy is put between
homesrv and port 8081.
Workaround: none. It is silent by design and the fallback is never spoken.
## Open-Meteo's empty body becomes 0°C [#676] {#weather}
Costs: he is told the weather is clear and 0°C when the service answered
nothing. Distinct from V-589, which covered the HTTP status and not the body.
Revisit when: a weather answer is reported as wrong, or the geocoder changes.
Workaround: none.
+27
View File
@@ -0,0 +1,27 @@
# Unguarded invariants
`CLAUDE.md` names these as load-bearing. Nothing enforces either one. A rule
that lives only in prose gets broken by whoever did not read the prose. Both of
these fail silently when broken.
`tokenizerRev` and `preRouteLadder` were checked and need nothing. The rev is
baked into the embedder key, so a bump triggers re-embedding. A missing ladder
rung is observable in the decision record.
## heads_path may equal model_path [#692] {#heads}
Costs: the routing heads then score with the same graph the resident e5-small
uses, and recall degrades. There is no error and no log line, so it reads as
ordinary drift rather than a misconfiguration.
Revisit when: `deploy/mavend.json` is edited by hand, or a fine-tuned heads
graph is swapped in.
Workaround: check the two keys by eye. That is the whole guard today.
## baselineGrammars is mirrored by hand [#693] {#grammars}
Costs: the eval fixture restates the stage 0 rule set in the daemon's order,
and its own comment says so. Three test files score against it. A grammar added
to `buildRouter` alone means every routing measurement scores a set nobody
runs. `CLAUDE.md` warns about this failure by name.
Revisit when: the next stage 0 grammar is added. That is when it bites.
Workaround: add to both lists, which is what the rule already says.
+25
View File
@@ -0,0 +1,25 @@
# Layering and dead surface
Neither entry breaks anything today. Both make a later change cost more than it
should, which is why they are low and not medium.
## Domain packages depend on store and IPC types [#685] {#dtos}
Costs: dialogue exposes `store.DialogueSessionRow` in its port, the pure
morning planner takes a `store.Fact`, and auth policy imports IPC method and
caller types. There is no Go import cycle. A schema change reaches further than
it should.
Revisit when: the dialogue or fact schema changes, or a second transport
appears beside IPC.
Workaround: none needed. It compiles and it is correct.
## Eleven symbols are unreachable [#686] {#deadcode}
Costs: extra API and test surface, and comments that claim callers which no
longer exist. Three of the eleven must not be deleted. `HisGender` is a
documented seam tied to V-399. `AudioDuration` duplicates `internal/audio` and
should call it. `CountWord` is a one-line alias nobody uses and can go.
Revisit when: `deadcode` is wired into the audit gate, which needs the
allowlist this entry describes. **An unannotated list invites deleting the
three above.**
Workaround: none needed.
+26
View File
@@ -0,0 +1,26 @@
# Security
Both entries are mavweb's passkey seam. Assertion itself is sound and is not
the problem: `userVerification` is required, and a sign count that does not
increase is rejected.
## Anyone past the proxy can enroll a passkey [#683] {#enrollment}
Costs: registration is gated on nothing, so any client that reaches mavweb can
enroll its own key and become him. Step-up is worse than per-client: one
process-global `assertedAt` means every client inherits the same five-minute
window after any successful assertion. The voice WebSocket accepts every
origin, which makes cross-site use easier. V-317 covers which routes are gated
and V-605 covers challenge-map growth. Neither covers this.
Revisit when: mavweb is reachable from anything but the tunnel, and before any
new credential is enrolled.
Workaround: the reverse proxy is the only boundary today. That is the finding.
## Passkey credentials are rewritten in place [#684] {#credentials}
Costs: `os.WriteFile` over the live file. A crash, a full disk or an
interrupted write corrupts every enrolled credential at once, and mavweb will
not start afterwards.
Revisit when: a second credential is enrolled, since the blast radius grows
with the count. Sooner if the box loses power unexpectedly.
Workaround: back the file up before enrolling.
+30
View File
@@ -0,0 +1,30 @@
# Storage
The DB seam: what it loses quietly, and what it reads more of than it needs.
## Dialogue persistence errors are swallowed [#677] {#dialogue}
Costs: restart continuity can vanish with nothing in the log, and a failed
delete can bring stale conversation state back. Current-turn dialogue is
unaffected, which is why this has never been noticed.
Revisit when: a restart is reported as losing context. Sooner if a turn starts
reading dialogue rows back to him.
Workaround: none. The failure is invisible from outside.
## The reminder transition is a lost update [#678] {#reminders}
Costs: a concurrent fire and cancel both succeed and the last writer wins.
Medium today because cancellation has no surface. High the moment V-622 adds
one, and V-622 does not describe this invariant.
Revisit when: V-622 starts, whichever comes first.
Workaround: none, but the window is small while nothing can cancel.
## A recall miss scans two whole tables [#681] {#recall}
Costs: every missed recall reads all of `memory_vectors` and then decodes and
sorts every note vector. Not an N+1, and the memory scan is cheap per losing
row on purpose. The duplicated decode is the legacy notes path alone.
Revisit when: the note count makes a miss measurably slow. Also when the two
exclusion filters are proven to agree. `QueryNotes` uses `notHisWordsSQL` and
`Search` uses `memory.NonRecallPrefix`. The fallback cannot go until they match.
Workaround: none needed at today's row counts.
+34
View File
@@ -0,0 +1,34 @@
# Transport
The HTTP and socket seams. What a client can do to them, and what a shutdown
can do to us.
## Closing a TCP listener can strand Accept [#679] {#accept}
Costs: during close, both `errc` and `done` are ready in `acceptLoop`'s select.
Go picks uniformly. So roughly one close in two leaves a waiting `Accept`
blocked forever on a TCP seam. Unix sockets are unaffected.
Revisit when: a daemon is seen hanging on shutdown, or before any new TCP
listener is added.
Workaround: the process usually exits anyway, which hides it.
## PTT reads an unbounded body [#688] {#ptt}
Costs: `handlePTT` does an unlimited `io.ReadAll`, and mavweb sets no header or
idle timeouts. A client can force unbounded allocation or hold a connection
open. mavgpud's half of this was fixed in V-673.
Revisit when: mavweb is reachable from anything but the tunnel.
Workaround: mavweb is not LAN-exposed today.
`/ws` rides along with this entry. It never calls `SetReadLimit`, so the
dependency default of 32,768 bytes applies, about a second of audio. Nothing
reaches it: the browser posts PCM to `/api/ptt`, and only `handlers_test.go`
opens `/ws`. It gets a caller and a real limit, or it gets deleted.
## mavweb errors cannot be traced [#689] {#errors}
Costs: some handlers return the raw internal error, which discloses internals.
Others return a generic one with no identifier, which cannot be joined to its
log line. There is no request-id middleware to join them.
Revisit when: a reported UI failure cannot be found in the log.
Workaround: read the log by timestamp.
+23
View File
@@ -0,0 +1,23 @@
# Background workers
Both entries are a tick doing expensive work it did not need to do.
## Fact enrichment is a 20-call serial waterfall [#680] {#enrichment}
Costs: each fact is resolved in turn and each ecosystem call can spend ten
seconds. A slow but reachable Nexus holds one tick for minutes, so the worker
stops observing its configured interval. V-647 covered the duplicate queue
scan, not this.
Revisit when: Nexus gets slow, or when a batch-resolution endpoint exists.
Workaround: an unreachable Nexus is fine. It is the slow-but-answering case
that hurts.
## A suppressed nudge is phrased anyway [#687] {#nudges}
Costs: `PhraseNudge` runs before the dedupe is known. The `continue` meant to
skip it is the last statement in the loop body. Every tick that
keeps suppressing the same rule pays the resident model again. The comment
above it claims the opposite.
Revisit when: digestion ticks show up in the model's load, or when nudge rules
grow past a handful.
Workaround: none.