diff --git a/docs/PORTING_NPE_PIPELINE.md b/docs/PORTING_NPE_PIPELINE.md
new file mode 100644
index 0000000..f1311c6
--- /dev/null
+++ b/docs/PORTING_NPE_PIPELINE.md
@@ -0,0 +1,123 @@
+# Porting NPE's player-JS pipeline into rustypipe
+
+**Branch:** `kayos/m1-sig-port`
+**Goal:** Replace `src/deobfuscate.rs`'s narrow regex approach with
+NewPipeExtractor's full pipeline so the fork keeps working as YouTube
+rotates its `player_ias.vflset/.../base.js`.
+
+## The diagnosis
+
+Upstream rustypipe 0.11.4 (June 2025) extracts the signature
+deobfuscation function with six regex patterns aimed at the call site
+(`var&&(var=SIGFN(decodeURIComponent(var)))`). On current YouTube player
+`c2f7551f` (May 2026) all six miss. NewPipeExtractor master's six
+patterns also miss on the same file — and NPE-master's nsig (throttling)
+pipeline is openly broken (`TeamNewPipe/NewPipeExtractor#1339`, open
+since 2026-02-03; the dev branch has had no sig/nsig commits in 60
+days). The reason NPE *appears* to work in apps is that the
+Innertube paths for Android / iOS / TV clients return stream URLs that
+don't carry an obfuscated `s=` signature for most videos — sig deobf
+is a fallback the typical playback path never reaches.
+
+Two structural changes have happened since rustypipe was last cut:
+
+1. **The sig fn call site now sometimes takes a numeric prefix arg.**
+   New shape: `var&&(var=SIGFN(123,decodeURIComponent(var)))`. NPE's
+   regex set has one pattern for this; rustypipe doesn't.
+
+2. **YT routes literal token references through a global string array.**
+   Near the top of every recent `player.js`:
+   ```js
+   var e="startsWith{redirector.googlevideo.com{split{...{decodeURIComponent{...".split("{")
+   ```
+   Calls then reference `e[N]` instead of the literal symbol. So an
+   anchor like `decodeURIComponent` is no longer present at the sig-fn
+   call site as text — it's `e[37]` (or whatever the index is).
+
+NPE's pipeline handles (1) but not (2). To make the fork robust we
+do both.
+
+## What we're porting
+
+| NPE file | Rust target | Notes |
+|---|---|---|
+| `YoutubeSignatureUtils.java` | `src/deobfuscate.rs` (rewritten) | Sig fn name + body + helper-obj + global-var assembly |
+| `YoutubeThrottlingParameterUtils.java` | new `src/deobfuscate/throttling.rs` module | nsig fn name + body + early-return fixup |
+| `utils/jsextractor/JavaScriptExtractor.matchToClosingBrace` | new `src/deobfuscate/jslexer.rs` | Find a `name=function` site, walk braces until balanced |
+| `YoutubeJavaScriptPlayerManager.java` | already covered by rustypipe's `cache.rs` | We keep rustypipe's cache shape but extend the cached payload to include nsig fn + global var |
+
+## Pipeline (the desired flow)
+
+```
+player.js (string)
+  │
+  ├── extract_sig_fn_name        // 6+ regex patterns, w/ globalVar[N] retry
+  │       │
+  │       └── fall back to:      // globalVar[N] indirection
+  │               1. extract_global_string_array_indices()
+  │               2. find N where arr[N] == "decodeURIComponent"
+  │               3. re-run patterns with `(?:decodeURIComponent|globalVar\[N\])`
+  │
+  ├── extract_sig_fn_body        // lexer brace-walk, regex fallback
+  ├── extract_global_var         // var X="...".split("{")  (verbatim)
+  ├── extract_helper_obj_name    // from inside fn body: [;,]NAME[..
+  ├── extract_helper_obj_body    // var NAME={...};
+  └── assemble:
+        globalVar + ";" + helperObj + ";" + deobfFn + ";" + callerFn
+        ── eval in rquickjs ──→ deobf_sig(input) ⇒ deobf(input)
+
+player.js (string)
+  │
+  ├── extract_nsig_fn_name       // 7 NPE patterns including arr-index variants
+  │       │
+  │       └── if array variant: resolve var NAME=[fn1,fn2,fnN]
+  │
+  ├── extract_nsig_fn_body       // lexer brace-walk
+  ├── fixup_early_return         // strip `if(typeof X==="undefined")return arg;`
+  └── eval in rquickjs ──→ deobf_nsig(input) ⇒ deobf(input)
+```
+
+## Milestones
+
+| ID | Subject | Effort | Gate |
+|---|---|---|---|
+| M1.1 | Port `matchToClosingBrace` (clean brace walker) to `src/deobfuscate/jslexer.rs` | S | Standalone unit test against a tiny `var Wka=function(d){return /,/}/` fixture |
+| M1.2 | Replace `get_sig_fn_name` with NPE's 6 patterns (including `(\d+,)decodeURIComponent`) | S | T-1 fixture is the prior-working `9216d1f7` player + new fixture `c2f7551f.js` |
+| M1.3 | Add `extract_global_string_array` returning `(var_name, Vec<String>)` | S | unit test for the `var e="…".split("{")` shape |
+| M1.4 | Add `extract_helper_obj_name` from fn body + `extract_helper_obj_body` | S | unit test against the `qB={w8:..,EC:..,Np:..}` style fixture |
+| M1.5 | Assemble globalVar + helperObj + sigFn + caller; round-trip via rquickjs | M | the existing `t_deobfuscate_sig` test fixture passes via new code path |
+| M1.6 | Add globalVar[N] indirection retry to sig fn name extraction | M | new test: a fixture where the call site uses `e[N]` instead of `decodeURIComponent` |
+| M1.7 | Port nsig pipeline (`YoutubeThrottlingParameterUtils`) — 7 patterns + array-resolution + early-return fixup | M | port + run NPE's `nsig_tests` table in `tests/sig_tests.rs` |
+| M1.8 | Add live integration test downloading current `player.js` and asserting round-trip end-to-end | S | `cargo test --features live -- t_update` |
+| M1.9 | Bump `Cargo.toml` to `0.12.0-sulkta.1`, tag, push to `Sulkta-Coop/rustypipe` `kayos/m1-sig-port` | S | clean release |
+
+## Not in M1 (parking lot)
+
+- Deno / external-JS-runtime swap (yt-dlp's path; we revisit if M1
+  doesn't hold).
+- Caching the assembled deobf code across processes (cookie-jar style
+  on Android).
+- N-tier fallback against multiple geo `player.js` variants if YT ever
+  splits them.
+
+## Why this is safe-ish to ship
+
+NPE's pipeline is what straw v0.1.0-X currently relies on for the rare
+videos that hit the sig path. Porting it 1:1 to Rust gives us a
+behavioural baseline equivalent to what NPE provides — no regression
+from the Java side. The globalVar[N] indirection added in M1.6 is the
+forward-looking piece that handles current `c2f7551f`-style
+obfuscation NPE doesn't yet handle. If M1.6 turns out unnecessary
+(e.g. NPE-dev lands its own fix first), we can pull the patterns into
+parity but keep our generalised resolution layer.
+
+## Tracking
+
+Workspace task IDs:
+- `#226` parent — fork + ship the patched fork
+- `#230` audit + port the sig pipeline (this milestone)
+- `#231` build pipeline + crafting-table integration
+
+When M1 lands, U-2..U-5 revival becomes a `Cargo.toml` dep flip in
+`rust/strawcore/` + cherry-pick of the parked commits
+(`7ff5ac79e..a13896f5e` on `Sulkta-Coop/straw`).