fix(math): restore priority-based comment routing (regression from #1961) - #2611

…ce (#2612)

Adds delphi/docs/MATH_ALGORITHM_HISTORY.md — a git-archaeology reference of the
core Polis math algorithms (comment routing, PCA/projection, extremity/repness,
clustering, consensus/moderation): introductions, numeric-output-changing edits,
and bugs, with verified commit hashes and dates. Includes the 2025 comment-routing
regression (#1961, fixed in #2611), the pre-2018 uniform-random routing history,
and how to map conversations to algorithm versions via the prod deploy timeline.

Sourced entirely from public git history and code — no production data.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

commit-id:31321c47

…gression reference

Adds `delphi/docs/MATH_ALGORITHM_HISTORY.md` — a git-archaeology reference of the core Polis math algorithms (comment routing, PCA/projection, extremity/repness, clustering, consensus/moderation): when each was introduced, every edit that changed numeric output, and known bugs, with verified commit hashes and dates. It includes the 2025 comment-routing regression (#1961, fixed in #2611), the pre-2018 uniform-random routing history, and a guide to mapping conversations to algorithm versions via the production deploy timeline.

Sourced entirely from public git history and code — no production data.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

commit-id:31321c47

…d) + Q2 prev-tick group-votes

## What

Two changes to comment prioritization:

- Restore the real `priority_metric` branching formula in BOTH engine modes. Clojure HEAD passes a real boolean since #2611 (conversation.clj:686, merged 2026-07-18), so the all-49 mirror of bug #1961 (the meta-tid routing bug that flattened every comment's priority to ~49, which #2571 had us reproduce for parity) is no longer the parity behavior — the correct formula is.
- Replicate quirk Q2 (`delphi/docs/CLOJURE_QUIRKS.md`): Clojure's `:comment-priorities` node SHADOWS its current-tick group-votes input with `(:group-votes conv)` — the PREVIOUS tick's stored value (conversation.clj:658). clojure-legacy mode (the engine setting that reproduces the old Clojure engine's behavior exactly) now consumes the captured previous-tick group-votes (`{}` on the first tick, matching Clojure's nil); improved mode keeps the current-tick read.

## How it works

The current tick's group-votes are stored on `self.group_votes` each tick in both modes (like `self.pca`), as the in-memory analogue of Clojure persisting `:group-votes` in `math_main` (the math results table). The restart seam — reloading group-votes on `from_dict` / poller reload — is deferred to the poller equivalence phase and noted in the journal.

Stale-mirror test sites reworked:

- The designed xfail on `test_priority_metric_non_meta_squared` is harvested: the formula test now gates.
- `TestD12CommentPriorities` now pins varied (non-all-49) Python output plus coverage, instead of the mirror signature.
- The exact-value comparison in `test_legacy_clojure_regression.py::test_comment_priorities` is xfailed for ALL variants — the reference blobs predate the Clojure #2611 fix, so value parity is validated by the H-B replay battery (the `scripts/certify.py` Clojure-vs-Python replay comparison) until the blobs are regenerated with a fixed generator.

## Testing

TDD: `tests/test_priority_unmirror.py` — RED observed (formula tests fail on the mirror; prev-tick flow + group_votes storage fail on the old wiring), 7 GREEN after. Targeted runs: 13 passed (unmirror + float-serialization), 21 passed (D12/priority subset), 4 xfailed (legacy regression value comparisons).

commit-id:a85047a7