chore: remove agent/claude/planning files from repo and update .gitignore (#320)

Co-authored-by: ralksta <ralksta@users.noreply.github.com>

.jules/bolt.md

@@ -1,23 +0,0 @@
-## YYYY-MM-DD - [Memoize grid item re-rendering]
-
-**Learning:** `useMemo` can significantly boost the performance of long list or grid components (like `PhotoGrid` component) that otherwise gets re-rendered on simple interactions like navigating the image lightbox. Running standard formatting `pnpm format` applies huge changes across entire directories, so it's generally best to format and clean up ONLY modified files to avoid huge diff pollution in code review.
-**Action:** Be extremely cautious of running format/lint commands at the root codebase; specify only modified directories or files, or avoid formatting if it cascades globally. Focus on targeted memoization where state changes dictate expensive array mapping operations.
-
-## 2024-05-28 - [Deduplicate concurrent API requests via Promise caching]
-
-**Learning:** When using `Promise.all` to fetch dependent data in parallel (e.g., getting multiple subpages and standalone albums that internally call a shared `getAlbums` function), the internal cache might not be populated in time. This leads to redundant concurrent network requests to the upstream server.
-**Action:** Implement Promise deduplication (request coalescing) by caching the pending Promise itself (e.g., in a `this.pendingPromise` class field) rather than just the final result. Return the pending Promise to subsequent callers until it resolves, ensuring only one network request is made.
-## 2024-05-29 - [Buffer inheritance optimization]
-
-**Learning:** `Buffer` inherits from `Uint8Array` in Node.js. Passing a `Buffer` into `new Uint8Array(...)` creates a completely new `ArrayBuffer` and copies all elements, which is an O(N) memory allocation and copy.
-**Action:** When decoding base64 strings or similar byte streams in Node.js where a `Uint8Array` is expected, return `Buffer.from(data, 'base64')` directly instead of wrapping it.
-## 2026-04-24 - [OOM prevention through promise chunking]
-**Learning:** When fetching unbounded large collections (like photo albums and their assets), using concurrent `Promise.all` over the entire array can cause Node.js Out of Memory (OOM) crashes.
-**Action:** Instead of `Promise.all(array.map(...))`, use a chunked data fetching approach (e.g., `Promise.all` within a `for` loop processing chunks of e.g. 10 items) to balance memory constraints with concurrent network speed.
-
-## 2024-05-30 - [O(1) Memory Bucketing for Coordinate Aggregation]
-**Learning:** When bucketing or grouping large datasets (e.g., thousands of map coordinates into geographical clusters), array allocations (`[].push()`) followed by `reduce` operations create an O(N) memory overhead per bucket and expensive post-processing loops.
-**Action:** Replace memory-heavy array collections with running statistical accumulators (e.g., `latSum`, `lngSum`, `count`) directly on the bucket. This reduces bucket memory from O(N) to O(1).
-## 2024-05-29 - Memoize repeated auth checks in Map API
-**Learning:** When filtering map locations, the API performs redundant `isAuthenticated` checks for albums within the same subpage slug. Since each check involves cryptographic operations (HMAC calculations), doing this repeatedly in a nested filter loop severely impacts performance for galleries with many geotagged photos.
-**Action:** Use a request-scoped `Map` to memoize the results of `isAuthenticated` calls within the route handler, preventing redundant cryptographic operations for the same subpage context.

.jules/sentinel.md

@@ -1,27 +0,0 @@
-## 2024-05-18 - Unprotected Dynamic Image Generation Endpoints
-
-**Vulnerability:** The `/api/og` route generates an `ImageResponse` dynamically using `next/og`. Because this operation is computationally expensive and memory-intensive, unauthenticated endpoints lacking rate limiting act as prominent Denial of Service (DoS) vectors. Attackers can flood the endpoint with varying parameters, forcing the server to continually spawn compute-heavy tasks until resources are exhausted and the instance crashes or latency spikes to unacceptable levels.
-
-**Learning:** Next.js dynamic endpoints that do not hit external upstream APIs or databases (like image generation with `next/og`) are often overlooked for rate limiting. Rate limiting is just as critical for protecting local compute resources as it is for protecting downstream API quotas or databases. Any route performing on-the-fly heavy processing must implement throttling.
-
-**Prevention:** Apply the shared `checkRateLimit` utility (from `@/lib/rate-limit`) to all computationally expensive endpoints (such as `next/og` usage), even if they do not explicitly query external services. Ensure `getConfig().rateLimitRpm` is passed as the threshold to maintain configurable global protection.
-## 2024-05-18 - [Preventing Timing Attacks on Variable-Length Secrets]
-**Vulnerability:** Comparing variable-length plaintext passwords using strict equality (`===`) or `crypto.timingSafeEqual` with unequal lengths, which leaks length information and enables timing attacks.
-**Learning:** When comparing variable-length secrets (like plaintext fallback passwords), both values must be hashed to a fixed-length algorithm (like SHA-256) first to ensure identical input lengths for the secure comparison, preventing side-channel leaks.
-**Prevention:** Always hash variable-length secrets to a fixed length before passing them to `crypto.timingSafeEqual`.
-## 2024-05-18 - Rate Limit Store Flooding and Cache Eviction Vulnerability
-**Vulnerability:** The rate limiter implemented an LRU/FIFO map to track API request limits, evicting the oldest tracked IP when maximum capacity was reached. An attacker could flood the store with randomly spoofed IP headers, continuously evicting legitimate users' IP addresses. This allowed the attacker to effectively bypass their own rate limit and subject the upstream server to DoS.
-**Learning:** Evicting entries from a security control store (like a rate limit map) under pressure turns the limit into a revolving door if an attacker controls the keys (e.g., via IP spoofing headers).
-**Prevention:** Rather than evicting old entries when an in-memory rate limit store is full, the correct security posture is to fail closed (or block new keys) until entries naturally expire, preserving tracking on existing keys.
-## 2026-04-23 - Prevent Global DoS in Rate Limiter via FIFO Eviction
-**Vulnerability:** The in-memory rate limiter implemented a hard ceiling (`MAX_STORE_ENTRIES`) and rejected all new IPs when reached, creating a trivial Global Denial of Service (DoS) vulnerability where an attacker could spam spoofed IPs to fill the bounded Map and permanently block all legitimate traffic.
-**Learning:** "Failing securely" by denying unknown requests during capacity limits is incorrect for rate limiters. Bounded Maps must actively evict older entries rather than rejecting new ones, otherwise the security mechanism itself becomes the vector for DoS.
-**Prevention:** In-memory rate limiters using bounded Maps must implement a FIFO (First-In-First-Out) or LRU (Least Recently Used) eviction policy (e.g., deleting `store.keys().next().value`) to make room for new IPs when maximum capacity is reached.
-## 2024-05-18 - Incorrect Scrypt Hash Generation in Plaintext Fallback
-**Vulnerability:** The plaintext password fallback logic incorrectly generated a recommended `scrypt` hash using the SHA256 hashed password rather than the raw `storedPassword`. This leads to generating an invalid `scrypt` hash that would fail authentication if the user updated their configuration with it. Furthermore, hashing variable-length secrets with SHA256 before `crypto.timingSafeEqual` prevents timing attacks, but using a non-keyed hash like SHA256 may leave it vulnerable. HMAC-SHA256 keyed with a secret provides better security.
-**Learning:** When generating a secure hash recommendation for users migrating from plaintext, ensure the raw input is used, not an intermediate hash intended for timing-safe comparison.
-**Prevention:** Generate the `scrypt` hash directly from the raw secret string (`storedPassword`). Always use HMAC keyed with a secret for hashing variable-length passwords prior to constant-time comparison.
-## 2024-05-30 - Prevent Stored XSS via Content-Type Validation on Proxied Assets
-**Vulnerability:** Asset proxy routes (`/api/image/*`, `/api/video/*`) directly served the upstream `Content-Type` header without validation. Since these API routes are explicitly excluded from the Next.js Content Security Policy (CSP) matcher, returning types like `image/svg+xml` or `text/html` would execute scripts directly in the user's browser, leading to Stored XSS.
-**Learning:** Endpoints that serve user-controlled content and are not protected by global CSP headers require strict MIME-type validation. An attacker could upload an SVG containing malicious scripts and retrieve it via these proxy endpoints to bypass standard security controls.
-**Prevention:** Always sanitize the `Content-Type` header on proxy endpoints. First convert it to lowercase, remap required generic types (like `application/octet-stream` to `image/jpeg`), and finally enforce a strict positive allowlist (`startsWith('image/')` and `!includes('svg')`) before falling back to `application/octet-stream`.