Rewrite AGENTS.md in short plain sentences (#327)

Every rule keeps its meaning and emphasis. Sentences are split so each
holds one idea, with the main rule before its exception and every
command, file name and identifier in backticks.

Three statements were false against the repo or the standing rulings
on record and are corrected:

- the format check is `make fmt-check`, not `make check-fmt`
- a feature branch reaches `next` through a PR that passes review, is
  merged only when the feature is completely finished and the code is
  up to standards, and is deleted after the merge; only the user
  merges `next` into `main`
- the naming example uses the code's term: `al` is bad,
  `SourceHostAllowlist` is good

The `Last Updated` line takes today's date.

Model: fable-5-1

AGENTS.md

@@ -1,75 +1,76 @@
 # Repository Rules
 
-Last Updated 2026-01-08
+Last Updated 2026-10-27
 
-These rules MUST be followed at all times, it is very important.
+These rules MUST be followed at all times. This is very important.
 
-- Never use `git add -A` - add specific changes to a deliberate commit. A commit
-  should contain one change. After each change, make a commit with a good
-  one-line summary.
+- Never use `git add -A`. Stage specific changes for a deliberate commit. A
+  commit holds one change. After each change, make a commit with a good one-line
+  summary.
 
 - NEVER modify the linter config without asking first.
 
-- NEVER modify tests to exclude special cases or otherwise get them to pass
-  without asking first. In almost all cases, the code should be changed, NOT the
-  tests. If you think the test needs to be changed, make your case for that and
-  ask for permission to proceed, then stop. You need explicit user approval to
-  modify existing tests. A bug fix may change an existing test without asking
-  when that test pins exactly the behavior the filed bug calls wrong, and the
-  change moves it to the fixed behavior and nothing else; any other change to an
-  existing test still needs explicit user approval. (You do not need user
-  approval for writing NEW tests.)
+- NEVER modify a test to exclude special cases or otherwise make it pass without
+  asking first. In almost all cases, change the code, NOT the test. You need
+  explicit user approval to modify an existing test. If you think a test needs
+  to change, make your case and ask for permission to proceed. Then stop. A bug
+  fix may change an existing test without asking only when that test pins
+  exactly the behavior the filed bug calls wrong. The change must then move the
+  test to the fixed behavior and nothing else. Any other change to an existing
+  test needs explicit user approval. Writing NEW tests needs no approval.
 
-- When linting, assume the linter config is CORRECT, and that each item output
-  by the linter is something that legitimately needs fixing in the code.
+- When linting, assume the linter config is CORRECT. Treat each item the linter
+  reports as something that legitimately needs fixing in the code.
 
 - When running tests, use `make test`.
 
-- Before commits, run `make check`. This runs `make lint` and `make test` and
-  `make check-fmt`. Any issues discovered MUST be resolved before committing
-  unless explicitly told otherwise.
+- Before committing, run `make check`. It runs `make test`, `make lint` and
+  `make fmt-check`. Any issue it finds MUST be resolved before committing,
+  unless you are explicitly told otherwise.
 
 - When fixing a bug, write a failing test for the bug FIRST. Add appropriate
-  logging to the test to ensure it is written correctly. Commit that. Then go
-  about fixing the bug until the test passes (without modifying the test
-  further). Then commit that.
+  logging to the test, to make sure the test is written correctly. Commit that.
+  Then work on the fix until the test passes, without changing the test further.
+  Then commit the fix.
 
-- When adding a new feature, do the same - implement a test first (TDD). It
-  doesn't have to be super complex. Commit the test, then commit the feature.
+- When adding a new feature, do the same: write a test first (test-driven
+  development, TDD). The test does not need to be very complex. Commit the test,
+  then commit the feature.
 
-- When adding a new feature, use a feature branch. When the feature is
-  completely finished and the code is up to standards (passes `make check`) then
-  and only then can the feature branch be merged into `main` and the branch
-  deleted.
+- When adding a new feature, use a feature branch. The feature branch reaches
+  `next` through a PR that passes review. It is merged only when the feature is
+  completely finished and the code is up to standards. The code is up to
+  standards when it passes `make check`. The branch is deleted after the merge.
+  Only the user merges `next` into `main`.
 
-- Write godoc documentation comments for all exported types and functions as you
-  go along.
+- Write `godoc` comments for all exported types and functions as you go along.
 
-- ALWAYS be consistent in naming. If you name something one thing in one place,
-  name it the EXACT SAME THING in another place.
+- ALWAYS be consistent in naming. If you name something one way in one place,
+  use the EXACT SAME name in every other place.
 
-- Be descriptive and specific in naming. `wl` is bad; `SourceHostWhitelist` is
+- Be descriptive and specific in naming. `al` is bad; `SourceHostAllowlist` is
   good. `ConnsPerHost` is bad; `MaxConnectionsPerHost` is good.
 
-- This is not prototype or teaching code - this is designed for production. Any
-  security issues (such as denial of service) or other web vulnerabilities are
-  P1 bugs and must be added to TODO.md at the top.
+- This is not prototype or teaching code. It is designed for production. Any
+  security issue (such as denial of service) or other web vulnerability is a P1
+  bug. It must be added at the top of `TODO.md`.
 
-- As this is production code, no stubbing of implementations unless specifically
-  instructed. We need working implementations.
+- Do not stub an implementation unless you are specifically instructed to. This
+  is production code, and we need working implementations.
 
 - NEVER silently fall back to a different setting when a user's parameter
-  explicitly specifies a value. If a user requests format=webp and WebP encoding
-  is not supported, return an error - do NOT silently output PNG instead. If a
-  user specifies fit=invalid and that fit mode doesn't exist, return an error -
-  do NOT silently default to "cover". Silent fallbacks violate the principle of
-  least surprise and mask bugs. The only acceptable defaults are for OMITTED
-  parameters, never for INVALID explicit values.
+  explicitly specifies a value. For example, if a user requests `format=webp`
+  and WebP encoding is not supported, return an error. Do NOT silently output
+  PNG instead. If a user specifies `fit=invalid` and that fit mode does not
+  exist, return an error. Do NOT silently default to `cover`. A silent fallback
+  violates the principle of least surprise and masks bugs. Defaults are
+  acceptable only for OMITTED parameters, never for INVALID explicit values.
 
-- Avoid vendoring deps unless specifically instructed to. NEVER commit the
-  vendor directory, NEVER commit compiled binaries. If these directories or
-  files exist, add them to .gitignore (and commit the .gitignore) if they are
-  not already in there. Keep the entire git repository (with history) small -
-  under 20MiB, unless you specifically must commit larger files (e.g. test
-  fixture example media files). Only OUR source code and immediately supporting
-  files (such as test examples) goes into the repo/history.
+- Avoid vendoring dependencies unless specifically instructed to. NEVER commit
+  the `vendor/` directory. NEVER commit compiled binaries. If such a directory
+  or file exists and `.gitignore` does not list it, add it to `.gitignore`.
+  Commit that `.gitignore` change. Keep the entire git repository, history
+  included, under 20 MiB. Go over 20 MiB only when you specifically must commit
+  a large file, such as a media file used as a test fixture. Only OUR source
+  code and its immediately supporting files, such as test examples, go into the
+  repository and its history.

README.md

@@ -133,7 +133,7 @@ What the [upaas](https://git.eeqj.de/sneak/upaas) app for pixa needs:
       filesystem's free space and what the cache holds, and at least 500 MiB, so
       the cache can grow to three quarters or more of the space they all have
       left, and to all of it when that is 500 MiB or less.
-    - the rest are in the table under Configuration below
+    - the rest are in the list under Configuration below
 - **Memory:** pixa can use up to `max_concurrent_processing` times 4.2 GiB for
   the images it processes at once, plus `upstream_connections` times
   `upstream_max_response_size` for the source images it has fetched that wait
@@ -178,47 +178,100 @@ allowlisted source hosts for open access.
 
 ## Design
 
+### Concepts
+
+These terms are the names the code uses:
+
+- **Source**: the image at one upstream URL, named by its host, path and query.
+- **Source key**: the SHA-256 of a source's host, path and query, which names
+  the source's file and its row in the database.
+- **Transform**: the output size, format, quality and fit an image request asks
+  for; with the format `auto`, pixa chooses the output format from the `Accept`
+  header.
+- **Image request**: a source and a transform, which each image route reads from
+  its URL.
+- **Variant**: the image pixa serves for an image request, made from its source
+  with its transform.
+- **Variant key**: the SHA-256 of a variant's source, size, format, quality and
+  fit, which names the variant's file and its row in the database, and gives its
+  `ETag`.
+- **Signed URL**: a `/v1/image/` URL whose `sig` signs its image request and its
+  expiry `exp` with `signing_key`.
+- **Encrypted URL**: a `/v1/e/` URL whose token holds its image request and any
+  expiry, encrypted with a key derived from `signing_key`.
+- **Login session**: an encrypted cookie that keeps a browser logged in to the
+  generator page for 30 days, or until a logout replaces the random value pixa
+  keeps in the `logout` table (see Encrypted URLs).
+- **Negative cache**: the source URLs pixa does not fetch again for 5 minutes,
+  after one of these failed fetches:
+    - the upstream answered with an error status
+    - the body's first bytes were not those of the format its `Content-Type`
+      names
+- **Image cache**: the part of pixa that keeps the sources, the variants and the
+  negative cache (see Storage).
+- **Disk cache limit** (`cache_max_bytes`): the size in bytes that the cached
+  sources and variants may reach together.
+- **Image proxy**: the part of pixa that serves a variant from the image cache
+  when it can, and otherwise makes it from the source.
+
 ### Storage
 
-- **Source images**:
-  `<state_dir>/cache/sources/<ab>/<cd>/<sha256 of host, path and query>`, one
-  file for each source URL
+- **Sources**: `<state_dir>/cache/sources/<ab>/<cd>/<source key>`, one file for
+  each source URL
 - **Database**: `<state_dir>/state.sqlite3` (SQLite)
-- **Transformed images**:
-  `<state_dir>/cache/variants/<ab>/<cd>/<sha256 of host, path, query, size, format, quality and fit>`
+- **Variants**: `<state_dir>/cache/variants/<ab>/<cd>/<variant key>`
 
 `<ab>` and `<cd>` are the first and second pairs of characters of the file's
 name.
 
-Each source image and each transformed image has one row in the database,
-holding its size and when it was last used. pixa reads a cached source image's
-format from its file's first bytes. When they match none of the source formats,
-pixa fetches the source again. Two source URLs that serve the same bytes are
-stored twice, once under each URL. A source image is stored by writing its file
-and then its row; when the row cannot be written, the file is removed again.
+The image cache keeps these tables in the database:
 
-Source and transformed image files are written under a temporary name, synced to
-disk, and only then renamed into place, so a crash leaves no file under its own
-name with part of its data. When pixa starts, and every five minutes after, it
-removes each cached source or transformed image whose file is empty, or not the
-size the database records for it, with a warning naming the file; the next
-request for it fetches the source or converts the image again. Looking for those
-files reads the database once for every cached file, so pixa skips each
-five-minute look that would start less than ten times as long after the last
-look ended as that look took. After a look that took one minute, the next one
-starts 10 to 15 minutes after that look ended.
+- `source_content`: a row for each source, by source key, with its size and when
+  it was last used
+- `variant_content`: a row for each variant, by variant key (column
+  `cache_key`), with its size and when it was last used
+- `cache_usage`: the total size of the cached sources and variants, which
+  eviction reads
+- `negative_cache`: a row for each source URL in the negative cache, with when
+  it expires
+
+pixa reads a cached source image's format from its file's first bytes. When they
+match none of the source formats, pixa fetches the source again. Two source URLs
+that serve the same bytes are stored twice, once under each URL. A source image
+is stored by writing its file and then its row; when the row cannot be written,
+the file is removed again.
+
+Each source and variant file is written under a temporary name and synced to
+disk before it is renamed into place. So a crash never leaves part of a file
+under its own name. When pixa starts, and every five minutes after, it removes
+each source or variant whose file is empty, or not the size its row records. It
+logs a warning naming the file. A later request for that image fetches the
+source or makes the variant again. Looking for those files reads the database
+once for every cached file, so pixa skips each five-minute look that would start
+less than ten times as long after the last look ended as that look took. After a
+look that took one minute, the next one starts 10 to 15 minutes after that look
+ended.
+
+Each look also drops the rows of variants whose files are missing. When pixa
+cannot read a variant file's size for another reason, it keeps the variant's
+row. It logs a warning naming the variant key and the error.
+
+During a look, pixa skips a directory it cannot list and a file whose size it
+cannot read. It logs a warning naming the path and the error, and goes on with
+the rest of the look. A file renamed or removed after pixa listed its directory
+gets no warning.
 
 pixa's target is 1-5k r/s, which has not been measured at that rate (see Load
-Test). Toward it, a cache hit opens the transformed image's file and nothing
-else, and makes no database call. The content type comes from the file's first
-bytes, which are those of the format pixa wrote it in; a file whose first bytes
-are not those of a format pixa writes is not served, and the image is converted
-again. A hit hands the time it was served to a background writer and does not
-wait for the writer. Eviction orders by that time. The hit is counted in memory.
-A miss that uses a cached source image also hands that writer the time it used
-the image. Once a second that writer writes the waiting times in one
-transaction. While 10,000 times of transformed images, or 10,000 of source
-images, wait to be written, a further one is dropped, and that image only looks
+Test). Toward it, a cache hit opens the variant's file and nothing else, and
+makes no database call. The content type comes from the file's first bytes,
+which are those of the format pixa wrote it in; a file whose first bytes are not
+those of a format pixa writes is not served, and the image is converted again. A
+hit hands the time it was served to a background writer and does not wait for
+the writer. Eviction orders by that time. The hit is counted in memory. A miss
+that uses a cached source image also hands that writer the time it used the
+image. Once a second that writer writes the waiting times in one transaction. Up
+to 10,000 access times of variants, and as many of sources, can wait to be
+written. A time handed over beyond that is dropped. Its image then only looks
 less recently used to eviction. `script/bench` measures how many hits pixa
 serves per second (see Load Test).
 
@@ -266,6 +319,8 @@ path under `/v1/` answers 200, in maintenance mode too.
     - 403 when the upstream host, or a host it redirects to, has an address in a
       blocked network (see `blocked_networks`).
     - 429 over `image_requests_per_minute` (below).
+    - 499, with no body, when the client goes away before its image is ready,
+      unless its request is the one making the image (below).
     - 415 when the upstream answers with a `Content-Type` that is not one of the
       source formats (below).
     - 415 when the upstream sends a body whose first bytes are not those of the
@@ -297,6 +352,7 @@ path under `/v1/` answers 200, in maintenance mode too.
         - 403
         - 415
         - 429
+        - 499, with no body
         - 502
         - 503
         - 504
@@ -678,54 +734,85 @@ refused with 400.
 Every setting can be given as an environment variable, in a YAML config file
 (`--config`), or both. A variable present in the environment wins over the file,
 even when it is empty, and the file wins over the built-in default. The one
-exception is a variable named in the file's `env:` section: it is set while the
-file loads, so it overrides both the environment the process was started with
-and the file's own key. A variable's value is parsed as the same text in the
-file would be. The three lists take comma-separated entries, with the spaces
-around each trimmed; an empty variable is an empty list. A value that does not
-parse or is invalid aborts startup, naming the variable. A variable whose name
-starts with `PIXA_` but is not in the table below, such as a misspelled one or
-`PIXA_PORT`, aborts startup naming it, as an unknown config key does. The one
-other accepted name is `PIXA_CONFIG_PATH`, the config file's path (like
-`--config`). The variables set by the file's `env:` section are checked the same
+exception is a variable set in the file's `env:` section. pixa sets that
+variable while loading the file, so the variable wins over the environment pixa
+started with. It also wins over the file's key for the same setting. A
+variable's value is parsed as the same text in the file would be. A setting that
+holds a list takes comma-separated entries, with the spaces around each trimmed.
+An empty variable is an empty list. A value that does not parse or is invalid
+aborts startup, naming the variable. pixa refuses to start when a variable
+starts with `PIXA_` but is not in the list below, such as a misspelled one or
+`PIXA_PORT`. The error names the variable, as it names an unknown config key.
+pixa also accepts `PIXA_CONFIG_PATH`, which names the config file, like
+`--config`. The variables set by the file's `env:` section are checked the same
 way.
 
-pixa reads at most one config file: the one given with `--config` (or `-c`),
-otherwise the one `PIXA_CONFIG_PATH` names, otherwise the first of these that
-pixa finds: `/etc/pixa/config.yml`, `/etc/pixa/config.yaml`,
-`~/.config/pixa/config.yml`, `~/.config/pixa/config.yaml`, then `config.yml` and
-`config.yaml` in the working directory. A named file that does not exist, cannot
-be read or does not parse aborts startup. Of the files pixa looks for on its
-own, only one that does not exist is passed over, without a message. One that
-pixa cannot read or parse aborts startup, naming the file. So does one in a
-directory pixa may not enter, whether or not it is there, since pixa cannot
-tell. With no file, pixa uses the environment and the defaults.
+pixa reads at most one config file. It reads the file given with `--config` (or
+`-c`), or else the one `PIXA_CONFIG_PATH` names. pixa refuses to start when that
+file does not exist, cannot be read or does not parse. Without `--config` or
+`PIXA_CONFIG_PATH`, pixa looks in `/etc/pixa/`, then in `~/.config/pixa/`, then
+in the working directory. In each directory it looks for `config.yml`, then
+`config.yaml`. It reads the first file it finds. It skips a missing file without
+a message. pixa refuses to start when a file it finds cannot be read or does not
+parse, and the error names the file. It also refuses to start when it may not
+enter one of these directories, whether or not the file is there, since it
+cannot tell. With no file, pixa uses the environment and the defaults.
 
-| Variable                             | Config key                      | Meaning                                                                      |
-| ------------------------------------ | ------------------------------- | ---------------------------------------------------------------------------- |
-| `PIXA_SIGNING_KEY`                   | `signing_key`                   | Required: secret for signed and encrypted URLs and login, 32+ characters     |
-| `PORT`                               | `port`                          | Port to listen on; default `8080`                                            |
-| `PIXA_STATE_DIR`                     | `state_dir`                     | Directory for the database and the disk cache; default `/var/lib/pixa`       |
-| `PIXA_DB_URL`                        | `db_url`                        | SQLite database URL; default `state.sqlite3` in the state directory          |
-| `PIXA_CACHE_MAX_BYTES`               | `cache_max_bytes`               | Disk cache limit in bytes; `0` disables it; default 75% of (free + cached)   |
-| `PIXA_ALLOWLIST_HOSTS`               | `allowlist_hosts`               | Upstream hosts served without a signature                                    |
-| `PIXA_REFERER_BLOCKLIST`             | `referer_blocklist`             | Hosts whose pages the image routes refuse with 403, by `Referer`             |
-| `PIXA_BLOCKED_NETWORKS`              | `blocked_networks`              | CIDR ranges never fetched from, on top of the built-in ones                  |
-| `PIXA_TRUSTED_PROXIES`               | `trusted_proxies`               | CIDR ranges of proxies whose `X-Forwarded-For` is believed; default RFC 1918 |
-| `PIXA_ALLOW_HTTP`                    | `allow_http`                    | Allow plain-HTTP upstreams, for testing only; default `false`                |
-| `PIXA_UPSTREAM_CONNECTIONS_PER_HOST` | `upstream_connections_per_host` | Concurrent connections per upstream host; default `20`                       |
-| `PIXA_UPSTREAM_CONNECTIONS`          | `upstream_connections`          | Concurrent connections to all upstream hosts together; default `64`          |
-| `PIXA_MAX_CONCURRENT_PROCESSING`     | `max_concurrent_processing`     | Images processed at once; default half the CPUs, at least 1                  |
-| `PIXA_IMAGE_REQUESTS_PER_MINUTE`     | `image_requests_per_minute`     | Image requests per client per minute; `0` turns it off; default `600`        |
-| `PIXA_UPSTREAM_FETCH_TIMEOUT`        | `upstream_fetch_timeout`        | Time allowed for one fetch from an upstream host; default `30s`              |
-| `PIXA_UPSTREAM_MAX_RESPONSE_SIZE`    | `upstream_max_response_size`    | Largest upstream response accepted, in bytes; default 50 MiB                 |
-| `PIXA_DOWNSTREAM_TIMEOUT`            | `downstream_timeout`            | Time allowed for answering one client request; default `60s`                 |
-| `PIXA_ACCESS_CONTROL_ALLOW_ORIGIN`   | `access_control_allow_origin`   | CORS origin allowed to read image responses: `*` or one origin; default `*`  |
-| `PIXA_METRICS_USERNAME`              | `metrics.username`              | Username for `/metrics`, which is served only when both are set              |
-| `PIXA_METRICS_PASSWORD`              | `metrics.password`              | Password for `/metrics`; set together with the username                      |
-| `PIXA_SENTRY_DSN`                    | `sentry_dsn`                    | Sentry DSN for error reporting; empty disables it                            |
-| `PIXA_DEBUG`                         | `debug`                         | Debug logging and plain-HTTP local development; default `false`              |
-| `PIXA_MAINTENANCE_MODE`              | `maintenance_mode`              | Answer image requests with 503; the health check stays 200; default `false`  |
+Each item below is one setting, named by its config key:
+
+- `signing_key`: variable `PIXA_SIGNING_KEY`, required. The secret for signed
+  URLs, encrypted URLs and login. It must be 32 characters or longer.
+- `port`: variable `PORT`, default `8080`. The port to listen on.
+- `state_dir`: variable `PIXA_STATE_DIR`, default `/var/lib/pixa`. The directory
+  for the database and the disk cache.
+- `db_url`: variable `PIXA_DB_URL`, default `state.sqlite3` in the state
+  directory. The URL of the SQLite database.
+- `cache_max_bytes`: variable `PIXA_CACHE_MAX_BYTES`, default 75% of the sum of
+  the free space and what the cache holds. The disk cache limit in bytes. `0`
+  turns the disk cache off.
+- `allowlist_hosts`: variable `PIXA_ALLOWLIST_HOSTS`, default none. The upstream
+  hosts whose images pixa serves without a signature.
+- `referer_blocklist`: variable `PIXA_REFERER_BLOCKLIST`, default none. The
+  image routes refuse a request with 403 when its `Referer` header names one of
+  these hosts.
+- `blocked_networks`: variable `PIXA_BLOCKED_NETWORKS`, default none. The CIDR
+  ranges pixa never fetches from, on top of the built-in ones.
+- `trusted_proxies`: variable `PIXA_TRUSTED_PROXIES`, default the RFC 1918
+  ranges. The CIDR ranges of the proxies whose `X-Forwarded-For` header pixa
+  believes.
+- `allow_http`: variable `PIXA_ALLOW_HTTP`, default `false`. Allows plain-HTTP
+  upstream hosts. It is for testing only.
+- `upstream_connections_per_host`: variable
+  `PIXA_UPSTREAM_CONNECTIONS_PER_HOST`, default `20`. The most connections to
+  one upstream host at once.
+- `upstream_connections`: variable `PIXA_UPSTREAM_CONNECTIONS`, default `64`.
+  The most connections to all upstream hosts together at once.
+- `max_concurrent_processing`: variable `PIXA_MAX_CONCURRENT_PROCESSING`,
+  default half the CPUs, at least 1. The most images processed at once.
+- `image_requests_per_minute`: variable `PIXA_IMAGE_REQUESTS_PER_MINUTE`,
+  default `600`. How many image requests one client may send per minute. `0`
+  turns the limit off.
+- `upstream_fetch_timeout`: variable `PIXA_UPSTREAM_FETCH_TIMEOUT`, default
+  `30s`. The time allowed for one fetch from an upstream host.
+- `upstream_max_response_size`: variable `PIXA_UPSTREAM_MAX_RESPONSE_SIZE`,
+  default 50 MiB. The largest upstream response accepted, in bytes.
+- `downstream_timeout`: variable `PIXA_DOWNSTREAM_TIMEOUT`, default `60s`. The
+  time allowed for answering one client request.
+- `access_control_allow_origin`: variable `PIXA_ACCESS_CONTROL_ALLOW_ORIGIN`,
+  default `*`. The CORS origin allowed to read image responses. It is `*` or one
+  origin.
+- `metrics.username`: variable `PIXA_METRICS_USERNAME`, default empty. The
+  username for `/metrics`. pixa serves `/metrics` only when the username and the
+  password are both set.
+- `metrics.password`: variable `PIXA_METRICS_PASSWORD`, default empty. The
+  password for `/metrics`. Set it together with the username.
+- `sentry_dsn`: variable `PIXA_SENTRY_DSN`, default empty. The Sentry DSN for
+  error reporting. An empty DSN turns error reporting off.
+- `debug`: variable `PIXA_DEBUG`, default `false`. Turns on debug logging and
+  plain-HTTP local development.
+- `maintenance_mode`: variable `PIXA_MAINTENANCE_MODE`, default `false`. While
+  `true`, pixa answers image requests with 503. The health check still
+  answers 200.
 
 Key settings in more detail:
 
@@ -849,9 +936,25 @@ See `configs/config.example.yml` for all options with defaults.
 
 ### Architecture
 
-- **Dependency injection**: Uber fx. `internal/app` builds every part of pixad
-  from the config in one fx graph. Each setting's default is declared once, in
-  `internal/config`.
+- **Dependency injection**: Uber fx. `internal/app` puts every part of pixad in
+  one fx graph. `internal/providers` builds the image cache, the image proxy and
+  the other parts that serve images and logins. Each setting's default is
+  declared once, in `internal/config`.
+- **Packages**: these packages under `internal/` work with the concepts above:
+    - `handlers`: the routes, which read the image request from the URL, check
+      the signed or encrypted URL, and make the login session
+    - `imagerequest`: the image request, with the source key and the variant key
+    - `imageproxy`: the image proxy
+    - `imgcache`: the image cache, which keeps its sources in one `FileStore`
+      and its variants in another
+    - `httpfetcher`: fetching sources from upstream hosts
+    - `magic`: reading an image's format from its first bytes
+    - `imageprocessor`: making a variant from a source with libvips
+    - `signature`: checking the `sig` of a signed URL
+    - `encurl`: making and reading the tokens of encrypted URLs
+    - `seal`: the keys derived from `signing_key`, and the encryption of those
+      tokens
+    - `session`: the login session
 - **HTTP router**: go-chi
 - **Image processing**: govips (CGO wrapper for libvips)
 - **Database**: SQLite via modernc.org/sqlite

TODO.md

@@ -30,6 +30,42 @@ P2: security: per-origin rate limiting
 
 # Completed Steps
 
+- 2026-10-27 the reconciliation pass skips a directory it cannot list and a file
+  whose info it cannot read, and goes on (closes #343). It stopped before its
+  later checks at a directory it could not list, or at a source or variant file
+  whose info it could not read.
+    - It logs the warning "skipped unreadable entry in cache directory", with
+      the path and the error.
+    - A file renamed or removed after its directory was listed gets no warning.
+    - `reconcileFiles` reads each file's info once, before the temp-file check,
+      and hands it to `sweepStaleTempFile`, `reconcileVariantFile` or
+      `reconcileSourceFile`.
+    - A cancelled context still stops the walk with its error.
+- 2026-10-26 the reconciliation pass drops a variant's accounting row only when
+  its file is missing (closes #340). Any other error reading the file's size
+  keeps the row. It dropped the row on any error and logged the file as missing.
+    - The kept row gets the warning "kept accounting row for unreadable variant
+      file", with the variant key and the error.
+    - The pass checks for `ErrNotFound`, which `FileStore.Size` returns for a
+      missing file.
+- 2026-10-26 on both image routes, a request is answered 499 and logged at info
+  level when its client goes away before its image is ready, unless it is the
+  request making the image (closes #339). It was answered 500 and logged at
+  error level, as pixa's own failure.
+    - The line is "client went away before the image was ready". It carries the
+      `request_id`, the host and the path.
+    - The 499 has no body, since nobody reads it.
+    - `serveImage` answers it before calling `respondImageError`, which cannot
+      see the request's context.
+    - An error wrapping `context.Canceled` while the request's context is live
+      is still answered 500 and logged at error level.
+- 2026-10-26 `README.md`'s Design section starts with a Concepts list (closes
+  #321). Each concept gets one sentence, under the name the code gives it.
+    - Storage calls the transformed images variants, in shorter sentences.
+    - Storage's paths end in the source key and the variant key.
+    - Storage lists the image cache's tables in the database.
+    - Architecture names `internal/providers`, and lists the packages that work
+      with the concepts.
 - 2026-10-26 one table in `internal/config` lists every setting once (closes
   #320). `Config.settings` gives each setting's key, its environment variable
   and the function that loads it.
@@ -46,10 +82,11 @@ P2: security: per-origin rate limiting
       loaded.
     - When a config holds more than one invalid value, the error names the one
       that comes first in the table.
-    - A test checks that `README.md` and `configs/config.example.yml` name the
+    - `README.md` lists the settings one item each, where it had a table. A test
+      checks that this list and `configs/config.example.yml` name exactly the
       keys of the table.
+    - The key constants moved to the tests, their only users.
     - Removed:
-        - the key constants
         - `isKnownConfigKey`, `envVarNames`, `strictLoader` and `validate`
         - `validateAllowlistHostsValue` and `getStringSlice`, with the three
           tests of `getStringSlice`

internal/config/documented_settings_internal_test.go

@@ -8,11 +8,10 @@ import (
 	"testing"
 )
 
-// TestReadmeNamesEverySetting checks that the Configuration section of
-// README.md names every key of the settings table in backticks. It also
-// checks that every name in backticks there written like a key with an
-// underscore, such as state_dir, is a key of the table. The other keys,
-// such as port and metrics.username, are checked only the first way.
+// TestReadmeNamesEverySetting checks that the list of settings in
+// README.md names exactly the keys of the settings table, one item each.
+// The list is the first one in the Configuration section, and it ends at
+// the first blank line. Each item starts with its key in backticks.
 func TestReadmeNamesEverySetting(t *testing.T) {
 	t.Parallel()
 
@@ -30,28 +29,48 @@ func TestReadmeNamesEverySetting(t *testing.T) {
 	section, _, _ = strings.Cut(section, "\n## ")
 	section, _, _ = strings.Cut(section, "\n### ")
 
+	_, list, found := strings.Cut(section, "\n- ")
+	if !found {
+		t.Fatal("README.md's Configuration section has no list")
+	}
+
+	list, _, _ = strings.Cut("- "+list, "\n\n")
+
+	itemKey := regexp.MustCompile("^- `([^`]+)`")
 	named := map[string]bool{}
 
-	for _, match := range regexp.MustCompile("`([^`]+)`").
-		FindAllStringSubmatch(section, -1) {
+	for line := range strings.SplitSeq(list, "\n") {
+		// A line that does not start with "- " goes on with the item above.
+		if !strings.HasPrefix(line, "- ") {
+			continue
+		}
+
+		match := itemKey.FindStringSubmatch(line)
+		if match == nil {
+			t.Errorf("README.md's list of settings has an item "+
+				"that does not start with a key in backticks: %s", line)
+
+			continue
+		}
+
+		if named[match[1]] {
+			t.Errorf("README.md's list of settings names %s twice", match[1])
+		}
+
 		named[match[1]] = true
 	}
 
 	for _, s := range new(Config).settings() {
 		if !named[s.key] {
-			t.Errorf("README.md's Configuration section does not name %s", s.key)
+			t.Errorf("README.md's list of settings does not name %s", s.key)
 		}
 
 		delete(named, s.key)
 	}
 
-	keyWithUnderscore := regexp.MustCompile(`^[a-z]+(_[a-z]+)+$`)
-
-	for name := range named {
-		if keyWithUnderscore.MatchString(name) {
-			t.Errorf("README.md's Configuration section names %s, "+
-				"which is not a setting", name)
-		}
+	for key := range named {
+		t.Errorf("README.md's list of settings names %s, "+
+			"which is not a setting", key)
 	}
 }
 

internal/handlers/client_gone_internal_test.go

@@ -0,0 +1,177 @@
+package handlers
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"slices"
+	"testing"
+	"time"
+
+	"github.com/go-chi/chi/v5/middleware"
+	"sneak.berlin/go/pixa/internal/httpfetcher"
+	"sneak.berlin/go/pixa/internal/imagerequest"
+)
+
+// heldFetcher sends on started when a fetch begins, and holds the fetch until
+// release is closed. It then fails the fetch as an upstream error.
+type heldFetcher struct {
+	started chan struct{}
+	release chan struct{}
+}
+
+// Fetch sends on f.started, waits for f.release to be closed, and returns
+// httpfetcher.ErrUpstreamError.
+func (f heldFetcher) Fetch(
+	context.Context, imagerequest.Source,
+) (*httpfetcher.FetchResult, error) {
+	f.started <- struct{}{}
+
+	<-f.release
+
+	return nil, httpfetcher.ErrUpstreamError
+}
+
+// readLogLines decodes the JSON log lines in logs.
+func readLogLines(t *testing.T, logs *bytes.Buffer) []map[string]any {
+	t.Helper()
+
+	t.Logf("logged:\n%s", logs.String())
+
+	var lines []map[string]any
+
+	dec := json.NewDecoder(logs)
+	for dec.More() {
+		var line map[string]any
+
+		err := dec.Decode(&line)
+		if err != nil {
+			t.Fatalf("decoding log line: %v", err)
+		}
+
+		lines = append(lines, line)
+	}
+
+	return lines
+}
+
+// TestImageRoutes_ClientGoneWhileWaitingAnswers499 checks that both image
+// routes answer 499 to a request whose client goes away while it waits for an
+// image another request is making. A first request's fetch is held, so the
+// image is being made, and a second request for the same image waits for it
+// until its client goes away. The second request must be logged once at info
+// level with its request ID and the source host and path. Nothing may be
+// logged at error level.
+func TestImageRoutes_ClientGoneWhileWaitingAnswers499(t *testing.T) {
+	t.Parallel()
+
+	const requestID = "test-request-id"
+
+	for route, imageURL := range photoURLsByRoute() {
+		t.Run(route, func(t *testing.T) {
+			t.Parallel()
+
+			var logs bytes.Buffer
+
+			fetcher := heldFetcher{started: make(chan struct{}), release: make(chan struct{})}
+			h, routes := newRoutesFetchingThrough(t, fetcher,
+				slog.New(slog.NewJSONHandler(&logs, nil)))
+			target := imageURL(t, h)
+
+			firstDone := make(chan struct{})
+
+			go func() {
+				defer close(firstDone)
+
+				sendGet(t, routes, target)
+			}()
+
+			<-fetcher.started
+
+			ctx, cancel := context.WithCancel(context.WithValue(t.Context(),
+				middleware.RequestIDKey, requestID))
+			rec := httptest.NewRecorder()
+			secondDone := make(chan struct{})
+
+			go func() {
+				defer close(secondDone)
+
+				routes.ServeHTTP(rec, httptest.NewRequestWithContext(
+					ctx, http.MethodGet, target, nil))
+			}()
+
+			// Give the second request time to start waiting for the first
+			// request's image. The fetch is still held, so the second request
+			// can end only because its client went away
+			time.Sleep(100 * time.Millisecond)
+			cancel()
+			<-secondDone
+
+			close(fetcher.release)
+			<-firstDone
+
+			if rec.Code != 499 {
+				t.Errorf("status = %d, want 499", rec.Code)
+			}
+
+			var clientGoneLines int
+
+			for _, line := range readLogLines(t, &logs) {
+				if line["level"] == slog.LevelError.String() {
+					t.Errorf("logged at error level: %v", line["msg"])
+				}
+
+				if line["level"] == slog.LevelInfo.String() &&
+					line["msg"] == "client went away before the image was ready" &&
+					line["request_id"] == requestID &&
+					line["host"] == signedHost &&
+					line["path"] == photoPath {
+					clientGoneLines++
+				}
+			}
+
+			if clientGoneLines != 1 {
+				t.Errorf("info line naming the request logged %d times, want 1",
+					clientGoneLines)
+			}
+		})
+	}
+}
+
+// TestImageRoutes_CanceledWithLiveRequestAnswers500 checks that both image
+// routes answer 500, logged as "failed to get image" at error level, for an
+// image service error that wraps context.Canceled while the request's own
+// context is live. Its client has not gone away, so the cancellation is
+// pixa's own failure.
+func TestImageRoutes_CanceledWithLiveRequestAnswers500(t *testing.T) {
+	t.Parallel()
+
+	for route, imageURL := range photoURLsByRoute() {
+		t.Run(route, func(t *testing.T) {
+			t.Parallel()
+
+			var logs bytes.Buffer
+
+			h, routes := newRoutesFetchingThrough(t,
+				failingFetcher{err: context.Canceled},
+				slog.New(slog.NewJSONHandler(&logs, nil)))
+
+			rec := sendGet(t, routes, imageURL(t, h))
+			if rec.Code != http.StatusInternalServerError {
+				t.Errorf("status = %d, want %d",
+					rec.Code, http.StatusInternalServerError)
+			}
+
+			if !slices.ContainsFunc(readLogLines(t, &logs),
+				func(line map[string]any) bool {
+					return line["level"] == slog.LevelError.String() &&
+						line["msg"] == "failed to get image"
+				}) {
+				t.Error(`no "failed to get image" line logged at error level`)
+			}
+		})
+	}
+}

internal/handlers/image.go

@@ -1,6 +1,7 @@
 package handlers
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	"io"
@@ -19,6 +20,12 @@ import (
 	"sneak.berlin/go/pixa/internal/magic"
 )
 
+// statusClientClosedRequest is the status nginx calls "client closed
+// request". The image routes answer it when the client goes away before its
+// image is ready, unless its request is the one making the image. Nobody reads
+// it; it is for logs and metrics.
+const statusClientClosedRequest = 499
+
 // HandleImage handles the main image proxy route:
 // /v1/image/<host>/<path>/<width>x<height>.<format>, or with no format
 // /v1/image/<host>/<path>/<width>x<height>
@@ -345,7 +352,8 @@ func notModified(w http.ResponseWriter, r *http.Request, etag string) bool {
 // serveImage serves the image that req asks for. Both image routes call it
 // once they have the request and its URL's expiry, which ends the
 // Cache-Control max-age. It chooses the format for the format auto, answers
-// If-None-Match with 304 and HEAD with the headers only, and answers an error
+// If-None-Match with 304 and HEAD with the headers only. It answers 499 when
+// the image service stopped because the client went away, and any other error
 // of the image service with respondImageError.
 func (s *Handlers) serveImage(
 	w http.ResponseWriter, r *http.Request,
@@ -362,6 +370,21 @@ func (s *Handlers) serveImage(
 	startTime := time.Now()
 
 	resp, err := s.imgSvc.Get(r.Context(), req)
+
+	// The client went away before its image was ready, which is not pixa's
+	// failure. A cancellation while the request's context is live is pixa's,
+	// and respondImageError answers it 500.
+	if errors.Is(err, context.Canceled) && r.Context().Err() != nil {
+		s.log.Info("client went away before the image was ready",
+			"request_id", middleware.GetReqID(r.Context()),
+			"host", req.Source.Host,
+			"path", req.Source.Path,
+		)
+		w.WriteHeader(statusClientClosedRequest)
+
+		return
+	}
+
 	if err != nil {
 		s.respondImageError(w, req, err)
 

internal/imgcache/eviction.go

@@ -515,12 +515,15 @@ func (c *Cache) reconcileAccounting(ctx context.Context) error {
 }
 
 // reconcileFiles walks baseDir, one of the cache's storage directories,
-// sweeping stale temp files and passing each other file, with its name, to
-// reconcileFile. Once ctx is cancelled it stops at the next file and
-// returns ctx's error.
+// sweeping stale temp files and passing each other file, with its info and
+// name, to reconcileFile. It skips a directory it cannot list and a file
+// whose info it cannot read, so one such entry does not stop the pass. It
+// logs a warning naming the path and the error. A file renamed or removed
+// after its directory was listed is skipped with no warning. Once ctx is
+// cancelled it stops at the next file and returns ctx's error.
 func (c *Cache) reconcileFiles(
 	ctx context.Context, baseDir string,
-	reconcileFile func(path string, entry fs.DirEntry, name string) error,
+	reconcileFile func(path string, info fs.FileInfo, name string) error,
 ) error {
 	return filepath.WalkDir(
 		baseDir,
@@ -529,19 +532,38 @@ func (c *Cache) reconcileFiles(
 				return ctx.Err()
 			}
 
-			if err != nil || entry.IsDir() {
-				return err
+			if err != nil {
+				c.log.Warn("skipped unreadable entry in cache directory",
+					"path", path, "error", err)
+
+				return nil
+			}
+
+			if entry.IsDir() {
+				return nil
+			}
+
+			info, err := entry.Info()
+			if errors.Is(err, fs.ErrNotExist) {
+				return nil
+			}
+
+			if err != nil {
+				c.log.Warn("skipped unreadable entry in cache directory",
+					"path", path, "error", err)
+
+				return nil
 			}
 
 			name := entry.Name()
 
 			if strings.HasPrefix(name, tempFilePrefix) {
-				c.sweepStaleTempFile(path, entry)
+				c.sweepStaleTempFile(path, info)
 
 				return nil
 			}
 
-			return reconcileFile(path, entry, name)
+			return reconcileFile(path, info, name)
 		},
 	)
 }
@@ -551,8 +573,8 @@ func (c *Cache) reconcileFiles(
 // and sweeping stale temp files.
 func (c *Cache) reconcileVariantFiles(ctx context.Context) error {
 	return c.reconcileFiles(ctx, c.variants.baseDir,
-		func(path string, entry fs.DirEntry, name string) error {
-			return c.reconcileVariantFile(ctx, path, entry, imagerequest.VariantKey(name))
+		func(path string, info fs.FileInfo, name string) error {
+			return c.reconcileVariantFile(ctx, path, info, imagerequest.VariantKey(name))
 		})
 }
 
@@ -562,16 +584,11 @@ func (c *Cache) reconcileVariantFiles(ctx context.Context) error {
 // crash or a restored snapshot can leave it, is removed with its row and
 // a warning naming it, so the variant is converted again.
 func (c *Cache) reconcileVariantFile(
-	ctx context.Context, path string, entry fs.DirEntry, cacheKey imagerequest.VariantKey,
+	ctx context.Context, path string, info fs.FileInfo, cacheKey imagerequest.VariantKey,
 ) error {
-	info, err := entry.Info()
-	if err != nil {
-		return fmt.Errorf("failed to stat variant file: %w", err)
-	}
-
 	var recordedSize int64
 
-	err = c.db.QueryRowContext(ctx,
+	err := c.db.QueryRowContext(ctx,
 		`SELECT size_bytes FROM variant_content WHERE cache_key = ?`, string(cacheKey),
 	).Scan(&recordedSize)
 
@@ -613,7 +630,9 @@ func (c *Cache) adoptVariantFile(
 }
 
 // reconcileVariantRows drops accounting rows whose variant files are
-// missing, so the database never references deleted content.
+// missing, so the database never references deleted content. A file whose
+// size cannot be read for another reason may still be on disk, so its row
+// stays, with a warning naming its key and the error.
 func (c *Cache) reconcileVariantRows(ctx context.Context) error {
 	var after imagerequest.VariantKey
 
@@ -641,6 +660,13 @@ func (c *Cache) reconcileVariantRows(ctx context.Context) error {
 				continue
 			}
 
+			if !errors.Is(err, ErrNotFound) {
+				c.log.Warn("kept accounting row for unreadable variant file",
+					"cache_key", key, "error", err)
+
+				continue
+			}
+
 			_, err = c.db.ExecContext(ctx,
 				`DELETE FROM variant_content WHERE cache_key = ?`, string(key))
 			if err != nil {
@@ -705,8 +731,8 @@ func queryStringColumn[T ~string](
 // and sweeping stale temp files.
 func (c *Cache) reconcileSourceFiles(ctx context.Context) error {
 	return c.reconcileFiles(ctx, c.srcContent.baseDir,
-		func(path string, entry fs.DirEntry, name string) error {
-			return c.reconcileSourceFile(ctx, path, entry, imagerequest.SourceKey(name))
+		func(path string, info fs.FileInfo, name string) error {
+			return c.reconcileSourceFile(ctx, path, info, imagerequest.SourceKey(name))
 		})
 }
 
@@ -717,16 +743,11 @@ func (c *Cache) reconcileSourceFiles(ctx context.Context) error {
 // is removed with its row and a warning naming it, so the source is
 // fetched again.
 func (c *Cache) reconcileSourceFile(
-	ctx context.Context, path string, entry fs.DirEntry, key imagerequest.SourceKey,
+	ctx context.Context, path string, info fs.FileInfo, key imagerequest.SourceKey,
 ) error {
-	info, err := entry.Info()
-	if err != nil {
-		return fmt.Errorf("failed to stat source content file: %w", err)
-	}
-
 	var recordedSize int64
 
-	err = c.db.QueryRowContext(ctx,
+	err := c.db.QueryRowContext(ctx,
 		`SELECT size_bytes FROM source_content WHERE source_key = ?`,
 		string(key),
 	).Scan(&recordedSize)
@@ -891,17 +912,12 @@ func (c *Cache) correctUsageTotal(
 
 // sweepStaleTempFile removes a temp file left behind by a crashed
 // write once it is old enough that no in-flight store can own it.
-func (c *Cache) sweepStaleTempFile(path string, entry fs.DirEntry) {
-	info, err := entry.Info()
-	if err != nil {
-		return
-	}
-
+func (c *Cache) sweepStaleTempFile(path string, info fs.FileInfo) {
 	if time.Since(info.ModTime()) < staleTempFileAge {
 		return
 	}
 
-	err = os.Remove(path)
+	err := os.Remove(path)
 	if err != nil && !os.IsNotExist(err) {
 		c.log.Warn("failed to remove stale temp file", "path", path, "error", err)
 

internal/imgcache/eviction_internal_test.go

@@ -9,8 +9,10 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
 	"sync"
+	"syscall"
 	"testing"
 	"time"
 
@@ -699,6 +701,291 @@ func TestStartEvictionReconcilesAccountingWithDisk(t *testing.T) {
 	}
 }
 
+// logLineNaming returns the first line of logs, written by slog's JSON
+// handler, that is at level and names the variant key, or "" if none is.
+func logLineNaming(logs, level string, key imagerequest.VariantKey) string {
+	for line := range strings.Lines(logs) {
+		if strings.Contains(line, `"level":"`+level+`"`) &&
+			strings.Contains(line, `"cache_key":"`+string(key)+`"`) {
+			return line
+		}
+	}
+
+	return ""
+}
+
+// TestReconcileVariantRowsDropsOnlyRowsOfMissingFiles gives
+// reconcileVariantRows two tracked variants. The first one's file is a
+// symbolic link to itself, so its size cannot be read. Its row must stay,
+// with a warning naming its key and the error. The second one has no file.
+// Its row must go, with the info line for a missing file.
+func TestReconcileVariantRowsDropsOnlyRowsOfMissingFiles(t *testing.T) {
+	t.Parallel()
+
+	cache, _ := newEvictionTestCache(t, 1<<30)
+	storeEvictionTestVariant(t, cache, testVariantKeyOne, []byte("variant one"))
+	storeEvictionTestVariant(t, cache, testVariantKeyTwo, []byte("variant two"))
+
+	unreadablePath := cache.variants.keyToPath(string(testVariantKeyOne))
+
+	err := os.Remove(unreadablePath)
+	if err != nil {
+		t.Fatalf("failed to remove the first variant's file: %v", err)
+	}
+
+	err = os.Symlink(unreadablePath, unreadablePath)
+	if err != nil {
+		t.Fatalf("failed to link the first variant's path to itself: %v", err)
+	}
+
+	_, statErr := os.Stat(unreadablePath)
+	if statErr == nil || errors.Is(statErr, fs.ErrNotExist) {
+		t.Fatalf("os.Stat() of the first variant: error = %v, "+
+			"want an error other than a missing file", statErr)
+	}
+
+	err = os.Remove(cache.variants.keyToPath(string(testVariantKeyTwo)))
+	if err != nil {
+		t.Fatalf("failed to remove the second variant's file: %v", err)
+	}
+
+	var logBuf bytes.Buffer
+
+	cache.log = slog.New(slog.NewJSONHandler(&logBuf, nil))
+
+	err = cache.reconcileVariantRows(t.Context())
+	t.Logf("reconcileVariantRows() error = %v, log: %s", err, logBuf.String())
+
+	if err != nil {
+		t.Fatalf("reconcileVariantRows() error = %v", err)
+	}
+
+	rowQuery := `SELECT COUNT(*) FROM variant_content WHERE cache_key = ?`
+
+	if n := countRows(t, cache, rowQuery, string(testVariantKeyOne)); n != 1 {
+		t.Errorf("the variant whose size cannot be read has %d rows, want 1", n)
+	}
+
+	if n := countRows(t, cache, rowQuery, string(testVariantKeyTwo)); n != 0 {
+		t.Errorf("the variant with no file has %d rows, want 0", n)
+	}
+
+	logs := logBuf.String()
+
+	warning := logLineNaming(logs, "WARN", testVariantKeyOne)
+	if !strings.Contains(warning, statErr.Error()) {
+		t.Errorf("no warning names %s and the error %q", testVariantKeyOne, statErr)
+	}
+
+	if line := logLineNaming(logs, "INFO", testVariantKeyOne); line != "" {
+		t.Errorf("the variant whose size cannot be read was logged at info: %s", line)
+	}
+
+	dropped := logLineNaming(logs, "INFO", testVariantKeyTwo)
+	if !strings.Contains(dropped, "dropped accounting row for missing variant file") {
+		t.Errorf("no info line says the file of %s is missing", testVariantKeyTwo)
+	}
+}
+
+// openDirectoryNearPathLimit makes a chain of directories inside dir, each
+// named with 255 zeros, the longest name Linux allows. Linux refuses a path
+// of 4096 bytes or more, and the chain goes as deep as it can below that.
+// It returns the chain's deepest directory as an os.Root, closed when the
+// test ends. An entry made through it with a 255-byte name has a path even
+// the root user cannot read. The os.Root can make that entry because it
+// names the entry only by its name, not by the whole path.
+func openDirectoryNearPathLimit(t *testing.T, dir string) *os.Root {
+	t.Helper()
+
+	name := strings.Repeat("0", 255)
+	deepest := dir
+
+	for len(filepath.Join(deepest, name)) < 4096 {
+		deepest = filepath.Join(deepest, name)
+	}
+
+	err := os.MkdirAll(deepest, StorageDirPerm)
+	if err != nil {
+		t.Fatalf("failed to make a chain of directories: %v", err)
+	}
+
+	root, err := os.OpenRoot(deepest)
+	if err != nil {
+		t.Fatalf("os.OpenRoot(%s) error = %v", deepest, err)
+	}
+
+	t.Cleanup(func() { _ = root.Close() })
+
+	return root
+}
+
+// makeEntriesTooLongToRead makes a new directory and a new file in the
+// deepest directory of the chain that openDirectoryNearPathLimit makes
+// inside dir. Even the root user can then neither list the new directory
+// nor read the new file's info. It returns the paths of the new directory
+// and the new file.
+func makeEntriesTooLongToRead(t *testing.T, dir string) []string {
+	t.Helper()
+
+	deepest := openDirectoryNearPathLimit(t, dir)
+	dirName := strings.Repeat("0", 255)
+
+	err := deepest.Mkdir(dirName, StorageDirPerm)
+	if err != nil {
+		t.Fatalf("failed to make a directory too deep to list: %v", err)
+	}
+
+	fileName := strings.Repeat("1", 255)
+
+	err = deepest.WriteFile(fileName, []byte("unreadable"), StorageFilePerm)
+	if err != nil {
+		t.Fatalf("failed to make a file too deep to read: %v", err)
+	}
+
+	return []string{
+		filepath.Join(deepest.Name(), dirName),
+		filepath.Join(deepest.Name(), fileName),
+	}
+}
+
+// makeTempFileTooLongToRead makes a temp file in the deepest directory of
+// the chain that openDirectoryNearPathLimit makes inside dir. Even the
+// root user then cannot read the temp file's info. It returns the temp
+// file's path.
+func makeTempFileTooLongToRead(t *testing.T, dir string) string {
+	t.Helper()
+
+	deepest := openDirectoryNearPathLimit(t, dir)
+	name := tempFilePrefix + strings.Repeat("1", 255-len(tempFilePrefix))
+
+	err := deepest.WriteFile(name, []byte("partial"), StorageFilePerm)
+	if err != nil {
+		t.Fatalf("failed to make a temp file too deep to read: %v", err)
+	}
+
+	return filepath.Join(deepest.Name(), name)
+}
+
+// assertSkippedEntryWarning checks that logs hold the warning "skipped
+// unreadable entry in cache directory" naming path, with the error
+// ENAMETOOLONG for that path.
+func assertSkippedEntryWarning(t *testing.T, logs, path string) {
+	t.Helper()
+
+	for line := range strings.Lines(logs) {
+		if strings.Contains(line, `"level":"WARN"`) &&
+			strings.Contains(line, `"msg":"skipped unreadable entry in cache directory"`) &&
+			strings.Contains(line, `"path":"`+path+`"`) &&
+			strings.Contains(line, path+": "+syscall.ENAMETOOLONG.Error()) {
+			return
+		}
+	}
+
+	t.Errorf("no warning %q names %s with the error %q",
+		"skipped unreadable entry in cache directory", path, syscall.ENAMETOOLONG)
+}
+
+// TestReconciliationGoesOnPastEntriesItCannotRead puts, in each storage
+// directory, a directory the pass cannot list and a file whose info it
+// cannot read. The pass walks each directory in name order, so it reaches
+// them before the cache's files, whose keys start with a letter. It must
+// log a warning naming each one's path and the error, and still do the
+// rest of its work:
+//   - adopt a variant file that has no row
+//   - drop the row of a variant whose file is missing
+//   - remove a source file that has no row
+//   - correct a wrong total cache usage
+func TestReconciliationGoesOnPastEntriesItCannotRead(t *testing.T) {
+	t.Parallel()
+
+	cache, _ := newEvictionTestCache(t, 1<<30)
+	ctx := t.Context()
+
+	untrackedVariant := []byte("variant with no row")
+
+	err := cache.variants.Write(string(testVariantKeyOne), untrackedVariant)
+	if err != nil {
+		t.Fatalf("failed to write a variant file with no row: %v", err)
+	}
+
+	storeEvictionTestVariant(t, cache, testVariantKeyTwo, []byte("variant two"))
+
+	err = os.Remove(cache.variants.keyToPath(string(testVariantKeyTwo)))
+	if err != nil {
+		t.Fatalf("failed to remove the second variant's file: %v", err)
+	}
+
+	untrackedSourceKey := "ffeeddcc0001"
+
+	err = cache.srcContent.Write(untrackedSourceKey, []byte("source with no row"))
+	if err != nil {
+		t.Fatalf("failed to write a source file with no row: %v", err)
+	}
+
+	_, err = cache.db.ExecContext(ctx,
+		`UPDATE cache_usage SET total_size_bytes = 1 WHERE id = 1`)
+	if err != nil {
+		t.Fatalf("failed to set a wrong total: %v", err)
+	}
+
+	unreadable := slices.Concat(
+		makeEntriesTooLongToRead(t, cache.variants.baseDir),
+		makeEntriesTooLongToRead(t, cache.srcContent.baseDir),
+	)
+
+	logs := reconcileAsAtStartup(t, cache)
+
+	for _, path := range unreadable {
+		assertSkippedEntryWarning(t, logs, path)
+	}
+
+	rowQuery := `SELECT COUNT(*) FROM variant_content WHERE cache_key = ?`
+
+	if n := countRows(t, cache, rowQuery, string(testVariantKeyOne)); n != 1 {
+		t.Errorf("the variant file with no row got %d rows, want 1", n)
+	}
+
+	if n := countRows(t, cache, rowQuery, string(testVariantKeyTwo)); n != 0 {
+		t.Errorf("the variant with no file kept %d rows, want 0", n)
+	}
+
+	_, err = os.Stat(cache.srcContent.keyToPath(untrackedSourceKey))
+	if err == nil {
+		t.Error("the source file with no row is still on disk")
+	}
+
+	usage, err := cache.UsageBytes(ctx)
+	if err != nil {
+		t.Fatalf("UsageBytes() error = %v", err)
+	}
+
+	if usage != int64(len(untrackedVariant)) {
+		t.Errorf("UsageBytes() = %d, want %d, the size of the adopted variant",
+			usage, len(untrackedVariant))
+	}
+}
+
+// TestReconciliationWarnsOfTempFilesWhoseInfoItCannotRead puts, in each
+// storage directory, a temp file whose info the pass cannot read. The
+// pass must log a warning naming each one's path and the error, as it
+// does for any other file whose info it cannot read.
+func TestReconciliationWarnsOfTempFilesWhoseInfoItCannotRead(t *testing.T) {
+	t.Parallel()
+
+	cache, _ := newEvictionTestCache(t, 1<<30)
+
+	unreadable := []string{
+		makeTempFileTooLongToRead(t, cache.variants.baseDir),
+		makeTempFileTooLongToRead(t, cache.srcContent.baseDir),
+	}
+
+	logs := reconcileAsAtStartup(t, cache)
+
+	for _, path := range unreadable {
+		assertSkippedEntryWarning(t, logs, path)
+	}
+}
+
 // TestPeriodicReconciliationAdoptsFileThatAppearsAfterStartup proves
 // reconciliation is not a one-shot startup-only pass: it must also run
 // on the periodic ticker, so a variant file that lands on disk with no