refactor(forge): simplify auth gating (#139)

* docs: document S3 backend, auto-auth, and `:Pending done` command

Problem: The S3 backend had no `:Pending s3` entry in the COMMANDS
section, `:Pending auth` only mentioned Google, the `sync` config
field omitted `s3`, `_s3_sync_id` was missing from the data format
section, `:Pending done` was implemented but undocumented, and the
README lacked a features overview.

Solution: Add `:Pending s3` and `:Pending done` command docs, rewrite
`:Pending auth` to cover all backends and sub-actions, update config
and data format references, add `aws` CLI to requirements, and add a
Features section to `README.md`.

* feat(forge): add forge link parser and metadata fetcher

Problem: no way to associate tasks with GitHub, GitLab, or Codeberg
issues/PRs, or to track their remote state.

Solution: add `forge.lua` with shorthand (`gh:user/repo#42`) and full
URL parsing, async metadata fetching via `curl`, label formatting,
conceal pattern generation, token resolution, and `refresh()` for
state pull (closed/merged -> done).

* feat(config): add forge config defaults and `%l` eol specifier

Problem: no configuration surface for forge link rendering, icons,
issue format, or self-hosted instances.

Solution: add `pending.ForgeConfig` class with per-forge `token`,
`icon`, `issue_format`, and `instances` fields. Add `%l` to the
default `eol_format` so forge labels render in virtual text.

* feat(parse): extract forge refs from task body

Problem: `parse.body()` had no awareness of forge link tokens, so
`gh:user/repo#42` stayed in the description instead of metadata.

Solution: add `forge_ref` field to `pending.Metadata` and extend the
right-to-left token loop in `body()` to call `forge.parse_ref()` as
the final fallback before breaking.

* feat(diff): persist forge refs in store on write

Problem: forge refs parsed from buffer lines were discarded during
diff reconciliation and never stored in the JSON.

Solution: thread `forge_ref` through `parse_buffer` entries into
`diff.apply`, storing it in `task._extra._forge_ref` for both new
and existing tasks.

* feat(views): pass forge ref and cache to line metadata

Problem: `LineMeta` had no forge fields, so `buffer.lua` could not
render forge labels or apply forge-specific highlights.

Solution: add `forge_ref` and `forge_cache` fields to `LineMeta`,
populated from `task._extra` in both `category_view` and
`priority_view`.

* feat(buffer): render forge links as concealed text with eol virt text

Problem: forge tokens were visible as raw text with no virtual text
labels, and the eol separator logic collapsed all gaps when
non-adjacent specifiers were absent.

Solution: add forge conceal syntax patterns in `setup_syntax()`, add
`PendingForge`/`PendingForgeClosed` highlight groups, handle the
`%l` specifier in `build_eol_virt()`, fix separator collapsing to
buffer one separator between present segments, and change
`concealcursor` to `nc` (reveal in visual and insert mode).

* feat(complete): add forge shorthand omnifunc completions

Problem: no completion support for `gh:`, `gl:`, or `cb:` tokens,
requiring users to type owner/repo from memory.

Solution: extend `omnifunc` to detect `gh:`/`gl:`/`cb:` prefixes and
complete with `owner/repo#` candidates from existing forge refs in
the store.

* feat: trigger forge refresh on buffer open

Problem: forge metadata was never fetched, so virt text highlights
could not reflect remote issue/PR state.

Solution: call `forge.refresh()` in `M.open()` so metadata is
fetched once per `:Pending` invocation rather than on every render.

* test(forge): add forge parsing spec

Problem: no test coverage for forge link shorthand parsing, URL
parsing, label formatting, or API URL generation.

Solution: add `spec/forge_spec.lua` covering `_parse_shorthand`,
`parse_ref` for all three forges, full URL parsing including nested
GitLab groups, `format_label`, and `_api_url`.

* docs: document forge links feature

Problem: no user-facing documentation for forge link syntax,
configuration, or behavior.

Solution: add forge links section to `README.md` and `pending.txt`
covering shorthand/URL syntax, config options, virtual text
rendering, state pull, and auth resolution.

* feat(forge): add `find_refs()` inline token scanner

Problem: forge tokens were extracted by `parse.body()` which stripped
them from the description, making editing awkward and multi-ref lines
impossible.

Solution: add `find_refs(text)` that scans a string for all forge
tokens by whitespace tokenization, returning byte offsets and parsed
refs without modifying the input. Remove unused `conceal_patterns()`.

* refactor: move forge ref detection from `parse.body()` to `diff`

Problem: `parse.body()` stripped forge tokens from the description,
losing the raw text. This made inline overlay rendering impossible
since the token no longer existed in the buffer.

Solution: remove the `forge.parse_ref()` branch from `parse.body()`
and call `forge.find_refs()` in `diff.parse_buffer()` instead. The
description now retains forge tokens verbatim; `_extra._forge_ref`
is still populated from the first matched ref.

* feat(buffer): render forge links as inline conceal overlays

Problem: forge tokens were stripped from the buffer and shown as EOL
virtual text via `%l`. The token disappeared from the editable line,
and multi-ref tasks broke.

Solution: compute `forge_spans` in `views.lua` with byte offsets for
each forge token in the rendered line. In `apply_inline_row()`, place
extmarks with `conceal=''` and `virt_text_pos='inline'` to visually
replace each raw token with its formatted label. Clear stale
`forge_spans` on dirty rows to prevent `end_col` out-of-range errors
after edits like `dd`.

* fix(config): remove `%l` from default `eol_format`

Problem: forge links are now rendered inline, making the `%l` EOL
specifier redundant in the default format.

Solution: change default `eol_format` from `'%l  %c  %r  %d'` to
`'%c  %r  %d'`. The `%l` specifier remains functional for users who
explicitly set it.

* test(forge): update specs for inline forge refs

Problem: existing tests asserted that `parse.body()` stripped forge
tokens from the description and populated `meta.forge_ref`. The
`conceal_patterns` test referenced a removed function.

Solution: update `parse.body` integration tests to assert tokens stay
in the description. Add `find_refs()` tests covering single/multiple
refs, URLs, byte offsets, and empty cases. Remove `conceal_patterns`
test. Update diff tests to assert description includes the token.

* docs: update forge links for inline overlay rendering

Problem: documentation described forge tokens as stripped from the
description and rendered via EOL `%l` specifier by default.

Solution: update forge links section to describe inline conceal
overlay rendering. Update default `eol_format` reference. Change
`issue_format` field description from "EOL label" to "inline overlay
label".

* ci: format

* refactor(forge): remove `%l` eol specifier, add `auto_close` config, fix icons

Problem: `%l` was dead code after inline overlays replaced EOL
rendering. Auto-close was always on with no opt-out. Forge icon
defaults were empty strings.

Solution: remove `%l` from the eol format parser and renderer. Add
`forge.auto_close` (default `false`) to gate state-pull. Set nerd
font icons: `` (GitHub), `` (GitLab), `` (Codeberg). Keep
conceal active in insert mode via `concealcursor = 'nic'`.

* fix(config): set correct nerd font icons for forge defaults

* refactor(forge): replace curl/token auth with CLI-native API calls

Problem: Forge metadata fetching required manual token management —
config fields, CLI token extraction, and curl with auth headers. Each
forge had a different auth path, and Codeberg had no CLI support at all.

Solution: Delete `get_token()` and `_api_url()`, replace with
`_api_args()` that builds `gh api`, `glab api`, or `tea api` arg
arrays. The CLIs handle auth internally. Add `warn_missing_cli` config
(default true) that warns once per forge per session on failure. Add
forge CLI checks to `:checkhealth`. Remove `token` from config/docs.

* refactor(forge): extract ForgeBackend class and registry

Problem: adding a new forge required touching 5 lookup tables
(`FORGE_HOSTS`, `FORGE_CLI`, `FORGE_AUTH_CMD`, `SHORTHAND_PREFIX`,
`_warned_forges`) and every branching site in `_api_args`,
`fetch_metadata`, and `parse_ref`.

Solution: introduce a `ForgeBackend` class with `parse_url`,
`api_args`, and `parse_state` methods, plus a `register()` /
`backends()` registry. New forges (Gitea, Forgejo) are a single
`register()` call via the `gitea_backend()` convenience constructor.

* ci: format

* fix(forge): fix ghost extmarks, false auth warnings, and needless API calls

Problem: extmarks ghosted after `cc`/undo on task lines, auth warnings
fired even when CLIs were authenticated, and `refresh()` hit forge APIs
on every buffer open regardless of `auto_close`.

Solution: add `invalidate = true` to all extmarks so Neovim cleans them
up on text deletion. Run `auth status` before warning to verify the CLI
is actually unauthenticated. Gate `refresh()` behind `auto_close` config.

* ci: typing and formatting

* refactor(forge): simplify auth gating and rename `gitea_backend`

Problem: forge auth/warning logic was scattered through
`fetch_metadata` — per-API-call auth status checks, `_warned` flags,
and `warn_missing_cli` conditionals on every fetch.

Solution: replace `_warned` with `_auth` (cached per session), add
`is_configured()` to skip unconfigured forges entirely, extract
`check_auth()` for one-time auth verification, and strip
`fetch_metadata` to a pure API caller returning `ForgeFetchError`.
Gate `refresh` and new `validate_refs` with both checks. Rename
`gitea_backend` to `gitea_forge`.

lua/pending/diff.lua

@@ -89,8 +89,7 @@ local function refs_equal(a, b)
   if not a or not b then
     return false
   end
-  return a.forge == b.forge and a.owner == b.owner
-    and a.repo == b.repo and a.number == b.number
+  return a.forge == b.forge and a.owner == b.owner and a.repo == b.repo and a.number == b.number
 end
 
 ---@param lines string[]