granthi-sync/README.md

# granthi-sync v1.2

The signup → download → link-folders → cloud product spine for the Granthi
forge, tested against the BETA forge (granthi-beta.shre.ai). Python 3 stdlib +
git CLI only — same portability heritage as the estate's `gitea_sync.py` mesh.

```
   ┌──────────────┐   device flow    ┌─────────────────┐
   │ granthi-sync │ ───────────────▶ │ shre-id Zitadel │
   │  (client,    │ ◀─────────────── │  id.shre.ai     │
   │  Mac/laptop) │   access token   └─────────────────┘
   │              │
   │              │  POST /v1/link {zitadel_access_token, device_name}
   │              │ ───────────────▶ ┌───────────────────────────────┐
   │              │ ◀─────────────── │ granthi-link  :3042           │
   │              │  {login, token}  │ (granthi VPS, tailnet-only)   │
   │              │                  │  · userinfo validation        │
   │              │  POST /v1/repos  │  · ensure Gitea user (admin)  │
   │              │ ───────────────▶ │  · mint scoped user token     │
   │              │                  └──────────────┬────────────────┘
   │              │   git push/fetch (user token    │ admin API
   │              │   via credential helper)        ▼
   │              │ ───────────────▶ ┌───────────────────────────────┐
   └──────────────┘                  │ BETA forge  :3041             │
                                     │ granthi-beta.shre.ai          │
                                     └───────────────────────────────┘
```

## Quickstart (invited user)

You need a shre-id account — an operator creates it; there is no open signup
(see "Invite-only story"). You also need to be on the tailnet: the
provisioning service is not publicly exposed yet.

```sh
git clone https://granthi.shre.ai/nirpa/granthi-sync.git
cd granthi-sync

# 1. Link this device. Prints a URL + code; approve it in a browser within
#    5 minutes. Creates your forge account and stores a scoped token in
#    ~/.granthi-sync/config.json (0600). You never see a forge password.
./bin/granthi-sync link

# 2. See what is already yours on the forge.
./bin/granthi-sync list

# 3. Either pull an existing repo down...
./bin/granthi-sync get <repo>           # or <owner>/<repo>, --into DIR

# 3b. ...or push a local folder up. It becomes a private repo.
./bin/granthi-sync add ~/work/notes

# 3c. ...or pull down everything this account is allowed to see.
./bin/granthi-sync get --all --into ~/granthi

# 4. Keep everything synced. Autocommits, ff-pulls, pushes; skips anything
#    that has diverged rather than merging or forcing.
./bin/granthi-sync watch                # --once for a single pass
./bin/granthi-sync status               # what is linked, mode, last sync

# 5. Go back to how a folder looked at some point in time.
./bin/granthi-sync snapshots ~/work/notes
./bin/granthi-sync restore  ~/work/notes --at 20260823T142530Z
```

Run `watch` as a background daemon on macOS with
`client/launchd/ai.granthi.sync.plist` (edit the script path, then
`launchctl bootstrap gui/$UID <plist>`).

**If the device code expires** (5 minutes, unapproved), nothing is created —
no account, no token, no partial state. Just run `link` again.

**Merging is a forge action, not a client one.** `watch` deliberately refuses
to merge; when a folder shows `DIVERGED` in `status`, resolve it in git or on
the forge web UI. The client will never force or auto-merge your work.

## Two modes, because two very different folders ask for this

A folder people sync is either *their documents* or *their git project*, and
the correct behaviour is opposite in each case. Each linked folder therefore
carries a `mode`.

| | `mirror` | `snapshot` |
|---|---|---|
| chosen for | a plain folder `add` turned into a repo | a folder that was already a git repo, and anything `get` clones |
| commits on your behalf | yes, `sync: <ISO ts>` | **never** |
| where work lands | the branch | `refs/granthi-backup/<device>/<ts>` |
| a restore point is | every commit | every snapshot |

`snapshot` mode is what "the work may not be committed, but it is still
backed up" means in git terms. Each pass loads a scratch index from HEAD,
stages the working tree into *that* index, writes a tree, and commits it with
`commit-tree`. HEAD, your index, your stash and every file on disk are
untouched — you can be mid-rebase with a dirty tree and the backup still
records exactly what is on the disk right now. The user's history stays the
user's.

Why a custom ref namespace: verified on the beta forge (Gitea 1.27.2) that
`refs/granthi-backup/...` is accepted, is readable through `ls-remote`, and
does **not** appear in the branch list. Under `refs/heads` a machine taking a
backup every 30 seconds would bury the branches a person actually made.

Snapshots are parented on HEAD and deliberately **not** chained to the
previous snapshot: chaining would keep every old snapshot reachable from the
newest, so pruning a ref would free nothing and retention would be
decorative.

**Retention** (or 30-second backups become a disk leak nobody can navigate):
everything is kept for 24 h, then thinned to hourly for 7 days, then daily.
Pruning runs at most hourly, per device, and only over that device's own
refs. A ref whose timestamp this version cannot parse is **kept** — deleting
the unrecognised is how a backup system loses the one thing someone needed.

**Restore never writes over the working tree.** `restore` materialises a
restore point into a *new* directory and refuses a non-empty destination.
Someone restoring a backup is already having a bad day; overwriting the files
they still have would make the recovery tool the second disaster.

## The credential helper must be the ONLY helper (found by live QA)

`credential.helper` is a list that accumulates across system, global and repo
config, and git asks every helper in it. A stock mac already has two —
`osxkeychain` from Xcode's gitconfig, and `store` from many people's
`~/.gitconfig` — and they lose in both directions:

* **reading:** a stale entry for the forge host answers before our helper, so
  pushes fail `remote: Failed to authenticate user` long after the token was
  rotated, and nothing in this tool's config explains why. This is exactly
  how the first live-QA run failed;
* **writing:** git calls `approve` on every helper after a successful auth,
  so `store` copies the forge token into `~/.git-credentials` **in
  plaintext**. Keeping the token in a 0600 file and out of remote URLs buys
  nothing if git then hands it to a plaintext store.

So `install_credential_helper` (and the `git clone` in `get`) sets an **empty**
`credential.helper` first, which resets the inherited list, then adds ours.
Exactly one helper serves this repo.

Corollary worth remembering: a token embedded in a remote URL gets saved by
`store` on first use. During QA a verification clone with a URL-embedded
token re-created the very entry that had just been cleaned out. That is the
whole reason this client passes tokens through a helper and never a URL.

## Device identity

`link` mints a uuid on first run and persists it in `~/.granthi-sync/config.json`
as `device_id`, and sends it to `/v1/link`. Hostnames are neither stable
(people rename laptops) nor unique (every new mac is "Mac mini"), so a
hostname cannot key a backup ref or a device registry — two machines would
overwrite each other's snapshots. The service-side device registry is the
next phase; the client leads so the id already exists when it lands.

## What "add the computer to the network" means — and does not

The onboarding shape is: download → login → **the device is federated to the
account** → the device can reach its repos.

The middle step is a *device registration*, not a network membership. Those
sound like one step and must not be built as one: this estate's tailnet is a
single flat private network carrying the granthi VPS, aros-vps, the Shadow
box and the Mac. Putting a customer's laptop on it to let them sync a folder
would hand that laptop L3 reach to every piece of infrastructure we run.

So:

* **Our own machines** may join the tailnet — that is an operator action with
  an operator's judgement behind it.
* **Customer devices never do.** Their transport is public HTTPS to
  granthi-link and the forge through cloudflared. That is the same exposure
  step already in the promotion window below, and it is what makes a genuinely
  new computer able to onboard itself at all — today `link` only works from
  inside the tailnet, which means "you can't set up a new computer without an
  operator first" is the honest status.

## Components

### `server/granthi_link.py` — provisioning service (granthi VPS)

* `GET /health`
* `POST /v1/link {zitadel_access_token, device_name}` → validates the token
  against `https://id.shre.ai/oidc/v1/userinfo`, applies the **identity
  binding rules** (below), mints a token scoped
  `write:repository,write:user`, returns
  `{gitea_base, login, token, token_name}`. POST bodies are capped at
  **64 KB** (413 beyond; missing `Content-Length` → 411, invalid → 400).
* `POST /v1/repos {token, name, private}` → creates the user repo with the
  USER token; clone/html URLs are rebased onto `public_gitea_base` because the
  container `ROOT_URL` (https://granthi-beta.shre.ai) does not resolve for
  tailnet-only clients.

Deployment: `/opt/granthi-link/{granthi_link.py,config.json,state.json}` +
systemd unit `granthi-link.service`; binds `127.0.0.1:3042` **and**
`100.111.127.127:3042` (tailnet). **Not publicly exposed** — see promotion
window. The service **refuses to start** (exit 2) unless `config.json` is
mode 0600/0400 and owned by the user it runs as — the config carries the
forge admin password, so permissive perms fail closed, not open.

#### Rate limiting

Sliding-window, per `(route, client)`, **on by default** — `/v1/link` creates
accounts and mints tokens, so unlimited has to be a deliberate config act,
never an omission. Defaults: `/v1/link` 5/hour, `/v1/repos` 60/hour. `GET
/health` is never limited. Over the limit → **429** with a `Retry-After`
header, decided *before* the body is read so an abusive caller costs nothing.

```json
"rate_limit": {
  "enabled": true,
  "trust_forwarded_for": false,
  "trusted_proxies": [],
  "rules": {"/v1/link": [5, 3600], "/v1/repos": [60, 3600]}
}
```

* Anything malformed **refuses startup** rather than silently meaning
  unlimited: a bad rule, a non-boolean `enabled`/`trust_forwarded_for` (JSON
  `null`, `0`, or the *string* `"false"` — every non-empty string is truthy),
  or a `rate_limit` that is not an object. `[0, N]` disables an endpoint
  outright.
* Denied requests are *not* recorded, so a client that keeps hammering cannot
  push its own window forward and lock itself out permanently.
* State is an in-process dict behind a lock — granthi-link is one
  `ThreadingHTTPServer`, so that is the entire store. **If this ever runs
  multi-process or multi-host, the limiter must move with it.** The clock is
  read *inside* the lock; taken outside, racing threads append out of order
  and both `Retry-After` and window reclamation silently go wrong.
* The key store is capped (`MAX_RATE_KEYS`) **per route**, not globally. At
  capacity it reclaims expired windows, and if every window is still live it
  **refuses the new key** — fail closed. Evicting a live window would let an
  attacker who can mint many distinct keys clear their *own* limit on demand.
  The per-route budget matters just as much: with one shared table, a flood
  of cheap `/v1/repos` keys would exhaust it and lock brand-new `/v1/link`
  clients out, turning the fail-closed guard into a cross-route DoS.
* Rule values must be real integers. `bool` subclasses `int` in Python, so
  `[5, true]` would otherwise pass as a **1-second** window — an hourly limit
  quietly becoming ~5/sec.
* `trust_forwarded_for` is **off** by default, and turning it on **requires a
  non-empty `trusted_proxies`** — the header is honored only when the socket
  peer is in that list. Without it, anyone reaching the origin directly (it
  also listens on the tailnet) could pick and rotate their own rate-limit key
  just by sending a header. Turn it on when exposing behind cloudflared,
  where every request otherwise arrives from the tunnel and one abuser would
  starve everyone. A caller can *prepend* anything to `X-Forwarded-For`; a
  trusted proxy *appends* the peer it actually saw, so the service reads the
  **last** entry, never the first, and requires it to parse as a real IP.
  `trusted_proxies` is validated at startup: it must be a list (a bare string
  would be iterated character by character), every entry a valid network, and
  wildcards (`0.0.0.0/0`, `::/0`) are refused outright — they would restore
  exactly the "trust anyone" hole the setting exists to close.

#### Identity binding (`state.json`)

`/v1/link` originally bound purely by `preferred_username` / email
local-part — any Zitadel identity whose derived login collided with an
existing account got a token for that account (account takeover). The
service now persists a map of Zitadel `sub` → Gitea login in
`/opt/granthi-link/state.json` (0600, atomic tmp+rename writes) and applies:

1. **Mapped sub** → always the mapped login, regardless of what the current
   userinfo claims. If the mapped login was deleted from the forge it is
   re-created only when the service created it originally; adopted accounts
   are refused (409).
2. **Unmapped sub, login free** → create the user, record the mapping
   (`created_by_service: true`). A concurrent-create 409 from Gitea is
   handled idempotently: the user is re-fetched and accepted only if its
   primary email is exactly the one this request would have set.
3. **Unmapped sub, login taken** → bind ONLY when the Gitea user's primary
   email equals the Zitadel userinfo `email` **and** `email_verified` is
   true (recorded with `created_by_service: false`); anything else →
   `409 login exists and is not linked to this identity`.

A Gitea token is **never minted before the binding rule passes**, and a
corrupt/unreadable `state.json` fails closed (500) instead of falling back
to an empty map.

**Migration (pre-1.1.0 accounts):** identities whose userinfo carries no
verified `email` (e.g. Zitadel *machine* users) cannot self-adopt an
existing forge login under rule (c) — for them the first link after the
upgrade would 409 forever. Any forge account the v1 service created before
this change must be seeded into `state.json` once, as
`{"<sub>": {"login": "<login>", "created_by_service": true, ...}}`, written
0600 atomically. As of the beta rollout the only such account is the E2E
machine user `granthi-sync-e2e` (seeded); the forge's human admin `nirpa`
is never provisioned through `/v1/link`, so nothing else needed seeding.

Empirically verified mechanics on Gitea **1.27.1**, re-probed and still
holding on **1.27.2** (beta forge, 2026-08-22 — all three results below
matched; the probe minted a token on the `granthi-sync-e2e` machine user and
deleted it again, `DELETE …/tokens/{id}` returning 204 under basic auth):

* Token minting: `POST /api/v1/users/{login}/tokens` returns **401 for
  token-authenticated sudo** (both `Sudo:` header and `?sudo=`); it only works
  with **admin basic auth + `Sudo: <login>` header** (201). The config
  therefore carries `admin_login`/`admin_password` (root-only, 0600) in
  addition to `admin_token` (minted via
  `gitea admin user generate-access-token`, used for all other admin calls).
* User creation: `source_id` is omitted → **local user** with a random
  30-char password, `must_change_password=false`, `visibility=private`.
  Rationale: `/api/v1/admin/identity-auth-sources` 404s on 1.27.1; the
  shre-id OAuth2 source is ID 1 (via `gitea admin auth list`), but users
  attached to an OAuth2 source cannot basic-auth and admin-created users get
  no `external_login_user` row anyway — first OIDC web login links by email
  regardless of this choice.
* `test_mode` (config flag, **never in production**): allows `/v1/link` to
  accept `test_userinfo` in the body instead of a Zitadel round-trip, so E2E
  can exercise the ensure-user + mint path headlessly. It is honored **only
  when the service environment also sets `GRANTHI_LINK_ALLOW_TEST_MODE=1`**;
  a config flag without the env gate is logged as an ERROR and ignored.

### `client/granthi_sync_client.py` (+ `bin/granthi-sync`) — client daemon

* `link [--server URL] [--token TOK]` — Zitadel **device flow** (native app
  `granthi-sync-device`, client_id `386909715541590022`, project
  granthi-forge `386906525790109702`; grants: device_code + refresh_token):
  prints the verification URL + user code, polls the token endpoint
  (`authorization_pending`/`slow_down` handled), then calls `/v1/link`.
  `--token` skips the device flow with a ready Zitadel token (headless/dev).
  Result stored in `~/.granthi-sync/config.json` (0600).
* `list [pattern]` — every repo the linked token can see, with the local
  folder each is already synced to. `pattern` narrows the table by name
  (substring, or a glob like `work-*`), case-insensitively, against both
  `owner/name` and the bare name. Filtering is display-only: the set already
  came from the forge under this account's token. Reads
  `GET /api/v1/user/repos` on the forge
  **directly** with the scoped user token — no granthi-link round-trip, so
  the read path needs no service change. Pagination is followed to a short
  page; if the `FORGE_MAX_PAGES` guard trips, the output says the list is
  incomplete rather than letting a bounded page read as the whole set.
* `get --all [--into DIR] [--mode M]` — clone every repo this account can
  see, skipping the ones already linked here. **"Only the repos they are
  granted" needs no client-side permission logic**: `/api/v1/user/repos` is
  evaluated by the forge against this account's own scoped token, so the
  list *is* the grant. A client-side filter would be a second opinion about
  someone else's authorisation. One repo failing does not abandon the rest,
  and a truncated listing is reported loudly — `--all` must never quietly
  mean "the first 2000". `alice/notes` and `bob/notes` both want
  `<base>/notes`; the second is cloned to `<base>/bob-notes` and the clash is
  logged, because reporting it as "already present" would leave the user
  believing they had pulled both.
* `get <repo|owner/repo> [--into DIR] [--mode M]` — the download half of
  `add`. Defaults to `snapshot` mode unless the repo carries a
  `.granthi-sync.json` marker saying otherwise, so a plain synced folder
  behaves the same on the second machine while someone's real project is
  never autocommitted onto. Clones
  with `--origin granthi` (the remote name `watch` looks for) and
  `-c credential.helper=…` (the repo does not exist yet, so the helper
  cannot be installed first; git also persists it into the new config), then
  registers the folder in `config.json` with the same shape `add` writes —
  so a cloned repo is picked up by `watch` immediately. Refuses a non-empty
  destination. Branch is read with `symbolic-ref` (an empty repo has an
  unborn HEAD) and falls back to `main`.
* `add <folder> [--name N] [--private|--public] [--mode M] [--force]` —
  `git init -b main` if needed, creates the cloud repo via `/v1/repos`, adds
  remote `granthi`, initial commit + push. The token is delivered by a **git
  credential helper** (the client's hidden `git-credential` subcommand
  reading the 0600 config) — never embedded in the remote URL (estate rule).
  Pushes the folder's **current** branch, not a hardcoded `main`: an existing
  repo may sit on `master` or a feature branch, and publishing that work
  under the wrong name is not a cosmetic error.
  Two guards, because `add -A` takes whatever it is given: a starter
  `.gitignore` is seeded when the folder has none (an existing one is never
  touched — it is the user's), and a folder over 20 000 files / 512 MB is
  refused unless `--force`. The seeded ignore file covers `.env`, `*.key`,
  `*.pem`, `id_rsa` and friends, and it governs snapshots too — the scratch
  index honours `.gitignore` exactly as a normal commit does.
  The size guard measures what git *would* sync, ignore rules included
  (including the machine's global excludes), because its own advice is "add
  a .gitignore for what should not sync" and advice that changes nothing is
  worse than none. It asks git through a **throwaway git dir outside the
  folder**, so a refused `add` leaves no `.git` behind in a directory the
  user never agreed to turn into a repo.
* `watch [--interval 30] [--once]` — per folder, by mode. `mirror`:
  autocommit (`sync: <ISO ts>`) → fetch → ff-pull if remote strictly ahead →
  push if local strictly ahead. `snapshot`: fetch → push a snapshot of the
  working tree to this device's backup ref → ff-pull only when the tree is
  clean (local edits are already safe on the backup ref, so it reports and
  leaves the tree alone rather than failing) → push the user's own commits
  when they are strictly ahead. **DIVERGED → log + record + SKIP. Never
  force, never merge** — the same policy as the mesh — **but the backup
  still happens**, because divergence is when work is most at risk.
  Retention pruning runs at most hourly. SIGTERM-clean.
* `snapshots <folder> [--limit 20]` — restore points, newest first, **across
  every device**, with the device that took each one. Read from the
  **remote**, not a local cache: the feature exists for the case where this
  machine is gone.

  The three scopes differ deliberately. Writing is device-scoped, so two
  machines never overwrite each other. Pruning is device-scoped, so machine A
  never applies its clock to machine B's refs. **Reading is not scoped** — a
  replacement laptop has a new id, and scoping the read to it would print
  "no restore points yet" while the backups sit on the forge. That defect
  was live in the first draft and is now pinned by a test that restores a
  dead machine's work from a fresh clone.
* `restore <folder> --at <ts|sha> [--into DIR]` — materialise one restore
  point into a new directory; refuses a non-empty destination. Accepts what
  `snapshots` printed in either mode, including a mirror-mode `%cI`
  timestamp. Two commits inside the same second share that timestamp, so an
  ambiguous `--at` is **refused with the candidate ids** rather than
  resolved by guessing.
* `status` — table of linked folders, mode, last sync, divergence flags.
* Run as a daemon on macOS with `client/launchd/ai.granthi.sync.plist`
  (edit the script path, then `launchctl bootstrap gui/$UID <plist>`).

## Tests

* `python3 -m unittest discover -s tests` — 153 tests. The v1.2 additions
  cover: a snapshot capturing uncommitted work while HEAD, the index and the
  working tree stay byte-identical; snapshots landing outside `refs/heads`;
  an unchanged tree not being re-pushed; a diverged folder still being backed
  up; a dirty tree blocking the ff-pull but not the backup; retention keeping
  everything recent, thinning to hourly then daily, and **keeping**
  unparseable timestamps; prune deleting only the thinned refs; `.gitignore`
  seeding never overwriting an existing one and keeping `.env` out of
  snapshots; the folder-size guard being bounded rather than walking the
  disk; mode detection; `list` filtering; `get --all` skipping what is
  already present, defaulting to snapshot mode, and shouting about
  truncation; `restore` writing a new folder, refusing a non-empty
  destination, and leaving the working tree alone; and the credential helper
  being the only one the repo consults, proven by driving
  `git credential fill` against a deliberately poisoned outer helper.
* Earlier suite: autocommit/ff/diverged
  logic against real temp git repos (including "diverged never touches the
  remote"), config 0600 handling (including umask-proof creation and a
  no-chmod guard), credential-helper quoting/injection, mocked device-flow
  polling, the full `/v1/link` + `/v1/repos` service flows against an
  in-process stub playing Zitadel + Gitea, all identity-binding rules
  (collision 409, verified-email adoption, deleted-login re-create/refuse,
  concurrent-create race, corrupt-state fail-closed), the test_mode env
  gate, config-permission refusal, and the 64 KB body cap. The `list`/`get`
  set covers pagination-to-a-short-page, truncation being reported rather
  than hidden, HTTP errors being fatal instead of a silent empty list,
  non-empty-destination refusal, owner-qualified names, unborn-HEAD branch
  fallback, no token in the remote URL, and — the one that matters — that a
  `get` folder is actually picked up by a subsequent `sync_folder` pass.
* Live E2E against the beta forge is recorded in the delivery notes
  (link → add → watch ff/push → forced divergence → DIVERGED skip verified
  via API, remote sha untouched).

## Invite-only story (v1)

There is no open signup. An operator invites a user by creating them in
shre-id (Zitadel org). The user downloads the client, runs
`granthi-sync link`, signs in at id.shre.ai with the printed device code, and
the provisioning service creates their forge account + scoped token on the
fly — the forge never sees a password and the user never sees the forge admin.
Every linked folder becomes a private repo under their account.

## Next phase — invites and per-repo access (designed, not built)

Today `/v1/link` creates an account and every folder becomes a private repo
under it. What is missing is the multi-person case: an existing account
inviting somebody, and that person's device waking up with access to *some*
repos and not others.

Shape this should take, so the next session does not re-litigate it:

* **Where grants live: granthi-link's own store, not Zitadel orgs.** The
  estate's house pattern is app-side tenancy tables with the IdP only
  providing identity (see the AROS `tenants` / `tenant_members` split).
  Grants therefore sit beside `state.json`, and **Gitea is the enforcement
  point** — a grant is materialised as a repo collaborator or an org team
  membership, so the forge itself refuses unauthorised reads. Nothing in the
  client decides access, which is why `get --all` needs no permission logic.
* **Token scope does not change.** `write:repository,write:user` stays; per
  repo permission is collaborator/team state, not a token property.
* **`POST /v1/invite`** (account admin → new member): creates the shre-id
  user (Zitadel admin PAT, on aros-vps at
  `/opt/shre-id/deploy/secrets/shre_id_zitadel_pat`), records the intended
  grants, and returns an invite the person redeems by running `link`. Until
  they redeem it, nothing exists on the forge.
* **`POST /v1/grants`** (account admin): add/remove repo access for a member;
  applies the change to Gitea and records it. Removing a grant must also
  remove the collaborator — a grant store that drifts from the forge is
  worse than no store.
* Both endpoints are account-admin-only and rate-limited like `/v1/link`.
* The grant store inherits the same fragility already noted for the rate
  limiter: a flat JSON file behind an in-process lock, fine for one
  `ThreadingHTTPServer` and **not** fine the day this runs multi-process.

## Promotion window (beta → prod)

1. **Expose :3042** behind cloudflared (granthi.shre.ai vhost or
   link.granthi.shre.ai) — today it is tailnet-only by design.
2. **Swap forge base URLs** in `/opt/granthi-link/config.json`:
   `gitea_base` → prod forge, `public_gitea_base` →
   `https://granthi.shre.ai`; the client default server URL moves to the
   public endpoint.
3. The `granthi-web` OIDC app already lists the prod callback; the device
   app is host-independent. Rotate the beta admin token/password out of the
   config when pointing at prod (prod forge is READ-ONLY to this estate —
   promotion is an operator action, not an agent action).
4. ~~Add rate limiting / abuse controls before public exposure.~~ **DONE** —
   see "Rate limiting" above (5 links/hour per client by default). When
   exposing behind cloudflared, set `trust_forwarded_for: true` in the same
   change, or every request will look like the tunnel and one abuser will
   throttle everybody.
5. **Hardening checklist (must all hold before exposing):**
   - [ ] `config.json` is 0600 (or 0400) and owned by the service user —
         the service refuses to start otherwise; verify with
         `systemctl status granthi-link` after any config edit.
   - [ ] `test_mode` is absent from the production config **and**
         `GRANTHI_LINK_ALLOW_TEST_MODE` is not set in the unit environment.
   - [ ] `/opt/granthi-link/state.json` exists, is 0600, and is included in
         VPS backups — losing it orphans sub→login bindings (existing users
         would need verified-email re-adoption).
   - [ ] `GET /health` returns 200 on both binds after restart.
   - [ ] Spot-check the identity map: a repeat link for a known sub returns
         the same login; a colliding username with a different sub gets 409.

<!-- review-service live check 20260819T131430Z -->