PR-C: implement the locked security-review rulings

Closes out the remaining findings from @ibnu76's review of 58971c05, per the
rulings locked in SESSION-STATE. Builds on merged main (aa7d477 = PR-A #12 +
PR-D #13). Pre-genesis, so storage-layout and ABI changes are free.

Rulings implemented

S-01 (Low) — bucketIndexRay == 0 sentinel collision → honest init (Option 1).
The pooled tail bucket used 0 as a lazy "means RAY" sentinel. Sustained losses
decay the index toward zero; the moment a settlement wrote exactly 0, the next
read resurrected every wiped bucket at face value — insolvency (~1400 days at
deployed rates, ~250 at the 5e18 policy cap; S01DeployParams has the sweep).
Now: _bucketAdd initializes the index to RAY explicitly on first entry (safe:
a floor-guaranteed nonzero index means 0 ⇔ no member has ever entered, zero scaled shares), _bucketIndex returns storage verbatim, and _settleBucket
floors the index at 1 wei — a fully wiped bucket is dust-dead, exits stay live
via the full-exit branch, and no value can be conjured back. _projectBucket
was rewritten to project through the index with identical operations and the
same floor, so view == materialized to the wei even in the wipe regime
(a V.8 fidelity improvement over the old live-based formula).

S-03 (Low) — sMax tracker, layered per the ruling.
(i) Never snap down. The old code snapped sMax to the tracked leader the
instant the leader shrank — a 1-wei dust post could drag sMax to dust and
inflate every post's participation factor to the clamp. Decay (10%/epoch,
30-epoch catch-up cap) is now the sole descent, floored at the tracked
leader. (ii) Permissionless refreshSMax(postId). The irreducible poke for
lazy-accrual dormant posts: settles the post if an epoch has passed, then feeds
its true stored total to the tracker. It can only make sMax more honest, so
it is open to anyone; the ops worker will poke the largest posts each epoch
(companion app-repo patch). Emits SMaxRefreshed. (iii) Tracker widened
3 → 10 (TRACKED_POSTS); getTopPosts() still returns the top three pairs
for ABI stability.

S-04 (closed, ratified) + I.4 re-documented. ECONOMIC_INVARIANTS.md I.4
is rewritten to be true-about-this-mechanism: sMax ≥ every tracked total after
update; deviations from untracked dormant posts are closeable by anyone via the
poke; participation clamps at 1.0 bound the pre-poke worst case at the rMax
ceiling. The cross-post coupling through global sMax is ratified as intended
design in one sentence (code comment at the participation site + I.4).

S-05 (Info) — false rMax < RAY comments. Both bucket-math sites now state
the true bound: rMax scales with elapsed epochs, rBase can exceed RAY under
dormancy, and the gRay >= RAY wipe branch is live (correct total-loss
floor), not dead as the old comment claimed.

S-09 (Info) — decay-rate docstring. 9e17 (10%/day) was always the deployed
constant and is correct as the backstop; the docstring claiming 0.5%/day was
the bug. Docstring fixed; Deploy.s.sol now sets the value explicitly on the
dev path and require-pins it on every path, so drift fails the deploy loudly.

S-11 (Info) — StakeLot.entryEpoch removed. Stored, never read by any
settlement path. Struct field, write site, and the view tuple slot are gone:
getUserLotInfo returns (amount, weightedPosition, sideTotal, positionWeight). Interface updated; all in-tree destructurings mechanically
migrated (36 sites).

S-12 (Info) — _rescalePositions removed. Positions are recomputed as
midpoints (< total) after every mutation and settlement, so the rescale body
was unreachable outside the neutral branch, where it was a no-op. Function,
both call sites, and the PositionsRescaled event are gone; the _applyEpoch
clamp remains as the safety net.

S-13 (Info) — dead ZeroAddress branch in acceptGovernance removed.
When pending is address(0) no caller can equal it, so NotPendingGovernance
always fired first. Behavior identical; regression test pins the revert.

Behavior change to be aware of (intended)

Under never-snap-down, an exited whale's sMax peak no longer vanishes
instantly — it decays (bounded, ≤ ~90 epochs to burn 4 orders of magnitude,
faster with per-epoch pokes), transiently suppressing smaller posts' rates.
That trade is the point: instant snap-down was the manipulation lever.
S04SpecTestPoC::test_Q2 now documents the full arc: suppression exists →
pokes burn the peak down → rate recovers to the never-whaled rate within 2%.

Tests

• Reviewer PoCs for fixed findings converted to regression form (insolvency →
  solvency asserts for S-01; drag/snap-down asserts inverted for S-03 with the
  poke closing the documented pre-poke gap; S-11 rewritten as a 4-tuple
  queue-position regression). Untouched PoCs remain as evidence.
• New test/PrCRegressions.t.sol: 7 targeted regressions (never-snap-down;
  exact decay curve + leader floor; permissionless poke honesty; 10-slot
  tracker; wiped-bucket solvency across two settlements incl. member exit;
  fresh-bucket re-init; S-13 revert).
• Gate: 332 passed / 0 failed across 55 suites, reproduced twice — once in
  the working sandbox, once on a pristine reconstruction of merged main with
  only the shipped editors applied.
• Post-patch src/StakeEngine.sol sha256:
  4c6f8dc99a10a0dd4ec2d47ace930a123a3f13fb48d3eeedeea62add777f17a9.

Toolchain note (worth reading — cost us a phantom-bug hunt)

solc 0.8.33 under this profile caches in-frame block.timestamp after the
first read, so chains of vm.warp(block.timestamp + X) inside one test frame
collapse to a single warp (the warp applies to the env; the test's re-read is
stale — engine/handler frames always read fresh, so contracts and PR-D's
invariant handler are unaffected). New tests use vm.getBlockTimestamp() for
warp arithmetic. Existing PoCs are left as authored: they are empirically
anchored on this exact toolchain, and their _fresh() absolute-warp pattern
re-anchors each scenario correctly.

Follow-ups

• Fresh Fuji redeploy required (storage + ABI changed): sequence with the
  Track B mock deploy; app/deployments/fuji.json addresses rotate and the app
  repo needs the 4-tuple getUserLotInfo ABI.
• Companion app-repo patch: ops-worker keeper poking refreshSMax each epoch
  for the largest posts.
• Slither: code removal may strand stale baseline fingerprints; if the CI gate
  complains, re-run with --update (expected shrink only, no new findings).