feat: initial v0.1 scaffold (#1)

* feat: initial v0.1 scaffold

Public API surface: get/2,3, put/3,4, fetch/3,4, delete/2, delete_many/2,
invalidate/2, clear/1. ETS adapter with TTL sweep + soft LRU max_size,
pg-based invalidation transport, ETS-backed single-flight fetch dedup.

Adapter-owns-its-process model; State opaque to nova_cache. Time arithmetic
in ms throughout. Invalidation is best-effort eventual, TTL is the
correctness backstop. Per-cache invalidation => best_effort | ttl_only |
strict, with strict refusing to start without ttl_default.

Ships with Taure/erlang-ci v2 CI + release workflows, ex_doc guides for
getting-started, adapters, invalidation, and telemetry.

* ci: fix workflow input names + enable ct/hank/mutate

Author: Taure

.github/workflows/ci.yml

@@ -0,0 +1,34 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+    paths-ignore:
+      - '**.md'
+      - 'guides/**'
+      - 'docs/**'
+      - 'LICENSE'
+  pull_request:
+    branches: [main]
+    paths-ignore:
+      - '**.md'
+      - 'guides/**'
+      - 'docs/**'
+      - 'LICENSE'
+
+jobs:
+  ci:
+    uses: Taure/erlang-ci/.github/workflows/ci.yml@v2
+    permissions:
+      contents: write
+      pull-requests: write
+    with:
+      enable-ct: true
+      enable-hank: true
+      enable-audit: true
+      enable-sbom: true
+      enable-sbom-scan: true
+      enable-dependency-submission: true
+      enable-mutate: true
+      enable-summary: true
+    secrets: inherit

.github/workflows/release.yml

@@ -0,0 +1,16 @@
+name: Release
+
+on:
+  push:
+    branches: [main]
+    paths-ignore:
+      - '**.md'
+      - 'guides/**'
+      - 'docs/**'
+      - 'LICENSE'
+
+jobs:
+  release:
+    uses: Taure/erlang-ci/.github/workflows/release.yml@v2
+    permissions:
+      contents: write

.gitignore

@@ -0,0 +1,12 @@
+_build/
+.rebar3/
+*.beam
+*.crashdump
+rebar.lock
+doc/
+ebin/
+.eunit/
+logs/
+*.iml
+.idea/
+.vscode/

CHANGELOG.md

@@ -0,0 +1,9 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+## [unreleased]
+
+### Features
+
+- Initial v0.1 scaffold: `nova_cache_adapter` and `nova_cache_invalidator` behaviours, public API, `nova_cache_ets` adapter, `nova_cache_invalidator_pg` transport, single-flight `fetch/3`.

README.md

@@ -1,2 +1,62 @@
 # nova_cache
-General-purpose KV cache library for the Nova ecosystem
+
+General-purpose KV cache library for the Nova ecosystem.
+
+`nova_cache` is **not** a dependency of Nova core and must never become one.
+
+## Quick start
+
+```erlang
+%% sys.config
+{nova_cache, [
+    {caches, #{
+        user_lookup => #{
+            adapter => nova_cache_ets,
+            ttl_default => 60_000,
+            max_size => 10_000
+        }
+    }}
+]}.
+
+%% application code
+ok = nova_cache:put(user_lookup, <<"alice">>, #{role => admin}),
+{ok, User} = nova_cache:get(user_lookup, <<"alice">>),
+{ok, User} = nova_cache:fetch(user_lookup, <<"alice">>, fun load_user/0).
+```
+
+## Adapters
+
+| Adapter            | Status |
+| ------------------ | ------ |
+| `nova_cache_ets`   | v0.1   |
+| `nova_cache_redis` | v0.2   |
+
+## Invalidation transports
+
+| Transport                   | Status |
+| --------------------------- | ------ |
+| `nova_cache_invalidator_pg` | v0.1   |
+
+## Build
+
+```sh
+rebar3 compile
+rebar3 dialyzer
+rebar3 xref
+```
+
+## Test
+
+```sh
+rebar3 ct
+rebar3 eunit
+rebar3 mutate
+```
+
+## Documentation
+
+See the [guides](guides/) directory.
+
+## License
+
+Apache-2.0.

cliff.toml

@@ -0,0 +1,38 @@
+[changelog]
+header = """
+# Changelog\n
+All notable changes to this project will be documented in this file.\n
+"""
+body = """
+{% if version %}\
+    ## [{{ version | trim_start_matches(pat="v") }}] - {{ timestamp | date(format="%Y-%m-%d") }}
+{% else %}\
+    ## [unreleased]
+{% endif %}\
+{% for group, commits in commits | group_by(attribute="group") %}
+    ### {{ group | striptags | trim | upper_first }}
+    {% for commit in commits %}
+        - {% if commit.scope %}*({{ commit.scope }})* {% endif %}\
+            {{ commit.message | upper_first }}\
+    {% endfor %}
+{% endfor %}\n
+"""
+trim = true
+
+[git]
+conventional_commits = true
+filter_unconventional = true
+split_commits = false
+commit_parsers = [
+    { message = "^feat", group = "Features" },
+    { message = "^fix", group = "Bug Fixes" },
+    { message = "^docs", group = "Documentation" },
+    { message = "^refactor", group = "Refactor" },
+    { message = "^test", group = "Testing" },
+    { message = "^chore\\(release\\)", skip = true },
+    { message = "^chore", group = "Miscellaneous" },
+    { message = "^ci", skip = true },
+]
+protect_breaking_commits = false
+tag_pattern = "v[0-9].*"
+sort_commits = "oldest"

guides/adapters.md

@@ -0,0 +1,38 @@
+# Adapters
+
+Adapters implement the `nova_cache_adapter` behaviour and own their own
+storage process. The `State` returned from `start_link/2`-time registration is
+opaque to `nova_cache` and passed back to every subsequent callback.
+
+## Shipped adapters
+
+### `nova_cache_ets`
+
+In-process ETS table per cache. Direct concurrent reads and writes from the
+caller's process. Periodic sweep purges expired rows. Soft `max_size`
+enforced at sweep time with LRU-on-table-order eviction.
+
+Configuration:
+
+| Option           | Default      | Notes                                     |
+| ---------------- | ------------ | ----------------------------------------- |
+| `ttl_default`    | `infinity`   | Default TTL applied when `put` omits one. |
+| `max_size`       | `infinity`   | Soft bound; evictions happen on sweep.    |
+| `sweep_interval` | `60_000`     | Milliseconds.                             |
+| `invalidation`   | `best_effort`| `best_effort | ttl_only | strict`.        |
+
+`strict` mode refuses to start without `ttl_default`.
+
+## Writing a new adapter
+
+1. `-behaviour(nova_cache_adapter).`
+2. Implement `start_link/2`, `get/2`, `put/4`, `delete/2`, `delete_many/2`, `clear/1`.
+3. Optionally implement `get_many/2` and `put_many/2`.
+4. Register with `nova_cache_registry:register(Name, ?MODULE, State)` from
+   `init/1` so the public API can route calls.
+5. Subscribe to invalidation events on startup if you want cluster
+   propagation.
+
+Adapter callbacks may execute in the caller's process or proxy through the
+adapter's own gen_server. That's the adapter's choice; the contract is the
+return values, not the process topology.

guides/getting-started.md

@@ -0,0 +1,71 @@
+# Getting Started
+
+## Installation
+
+```erlang
+{deps, [
+    {nova_cache, {git, "https://github.com/novaframework/nova_cache.git", {branch, "main"}}}
+]}.
+```
+
+## Configuration
+
+Declare your caches in `sys.config`:
+
+```erlang
+{nova_cache, [
+    {caches, #{
+        user_lookup => #{
+            adapter => nova_cache_ets,
+            ttl_default => 60_000,
+            max_size => 10_000,
+            sweep_interval => 60_000,
+            invalidation => best_effort
+        }
+    }},
+    {invalidator, nova_cache_invalidator_pg}
+]}.
+```
+
+One supervised process per declared cache starts under `nova_cache_sup`.
+
+## Reading and writing
+
+```erlang
+ok = nova_cache:put(user_lookup, <<"alice">>, User).
+{ok, User} = nova_cache:get(user_lookup, <<"alice">>).
+User = nova_cache:get(user_lookup, <<"alice">>, #{role => guest}).
+```
+
+## get-or-compute
+
+```erlang
+{ok, User} = nova_cache:fetch(user_lookup, <<"alice">>, fun() ->
+    case load_from_db(<<"alice">>) of
+        {ok, U} -> {ok, U};
+        not_found -> {error, not_found}
+    end
+end).
+```
+
+Concurrent callers for the same key are deduplicated via single-flight.
+Disable with `#{single_flight => false}` if you need pass-through semantics.
+
+## Negative caching
+
+Off by default. Opt in per call:
+
+```erlang
+%% short negative TTL (5 seconds):
+nova_cache:fetch(user_lookup, <<"alice">>, F, #{cache_errors => {true, #{ttl => 5_000}}}).
+```
+
+## Invalidation across the cluster
+
+```erlang
+ok = nova_cache:invalidate(user_lookup, <<"alice">>).
+```
+
+The configured invalidator transport broadcasts the event to every subscribing
+node. Each node purges its local copy on receipt. Delivery is best-effort
+eventual; see the Invalidation guide for the failure model.

guides/invalidation.md

@@ -0,0 +1,46 @@
+# Invalidation
+
+`nova_cache` ships cluster invalidation as a swappable transport behaviour.
+The default transport is `nova_cache_invalidator_pg`, built on `pg`.
+
+## Guarantee
+
+**Best-effort eventual.** TTL is the correctness backstop. A node that is
+netsplit, GC-paused, or just-joined may miss broadcasts and serve stale data
+until the row expires.
+
+## Per-cache mode
+
+Configured via the `invalidation` key in the cache spec:
+
+| Mode          | Behaviour                                                                |
+| ------------- | ------------------------------------------------------------------------ |
+| `best_effort` | Subscribe to broadcasts; serve stale on miss. Default.                   |
+| `ttl_only`    | Skip broadcasts entirely; rely solely on TTL.                            |
+| `strict`      | Best-effort plus refuses to start without `ttl_default`. Bounds staleness.|
+
+## Failure mode: a node misses a broadcast
+
+It serves stale data until the row's TTL elapses. If the row was written with
+`ttl => infinity`, it serves stale data indefinitely. This is the design.
+
+For workloads where "indefinitely stale" is unacceptable, use `strict` mode
+and a finite `ttl_default`.
+
+## Failure mode: a node joins late
+
+On join, the node's caches start empty. They subscribe and start receiving
+broadcasts immediately. There is no backfill of historical events. Existing
+entries on the joining node (e.g. after a netsplit heal) are not purged
+automatically in `best_effort` mode -- if you need that, run `clear/1` from
+your application's join handler.
+
+## Writing a new transport
+
+1. `-behaviour(nova_cache_invalidator).`
+2. Implement `start_link/1`, `subscribe/2`, `broadcast/2`.
+3. Deliver event payloads to subscribed handlers on every node.
+4. Set `{nova_cache, [{invalidator, your_module}]}` in `sys.config`.
+
+Transports must deliver events unchanged. They may drop events on failure
+without compromising correctness, because TTL is the backstop.

guides/telemetry.md

@@ -0,0 +1,36 @@
+# Telemetry
+
+`nova_cache` emits OpenTelemetry counters and spans when `opentelemetry_api`
+is available at runtime. The dependency is optional; callers without it still
+build and run.
+
+## Counters
+
+| Name                | Attributes              |
+| ------------------- | ----------------------- |
+| `nova_cache.hit`    | `cache.name`            |
+| `nova_cache.miss`   | `cache.name`            |
+| `nova_cache.evict`  | `cache.name`, `reason`  |
+
+`reason` is `ttl` (lazy expiry on get), `sweep` (periodic sweeper), or
+`max_size` (LRU eviction).
+
+## Spans
+
+| Name                  | Attributes                                  |
+| --------------------- | ------------------------------------------- |
+| `nova_cache.fetch`    | `cache.name`, `cache.key.length`, `result`  |
+| `nova_cache.put`      | `cache.name`, `cache.key.length`            |
+| `nova_cache.invalidate` | `cache.name`, `cache.invalidate.scope`    |
+
+`result` is `hit | miss | loaded | error`.
+
+## Enabling
+
+Add `opentelemetry_api` to your project's `rebar.config`. `nova_cache`
+detects the module at runtime and starts emitting events. No configuration on
+`nova_cache` itself is required.
+
+The OpenTelemetry-aware sibling library `opentelemetry_nova_cache` (planned
+for v0.2) will install the trace/metric pipeline; until then, configure it in
+your application's own OpenTelemetry setup.