Tick-gap count read as a percentage: 68% should be 42%, and it justifies a live constant

A tick-gap **count** was read as a **percentage** and has propagated from a research doc, through a
locked spec, into a live constant and a test comment.

## The arithmetic

`tools/combat/MECHANICS-VERIFICATION.md:169` records the histogram as counts:

> 161 exchange gaps, every one an integer multiple of 2.000s (68 x 1, 46 x 2, 35 x 3, 12 x 4, zero
> off-lattice)

`68 + 46 + 35 + 12 = 161`, matching the stated `n`. So one-tick gaps are `68/161` = **42%**, not 68%.

`tools/combat/archive/TICK-PHASE-REVIEW.md:165` reads it correctly ("42% of ticks"). Everything
downstream does not.

## Where the wrong figure now lives

| File | Text |
|---|---|
| `tools/combat/COMBAT-RAIL-SPEC.md:132` | "A one-tick gap is normal (68% of gaps)" — in a file headed **"Status: LOCKED. Implement as written."** |
| `mudsharp/Combat/ParticipantRoster.cs:66` | "One missed tick is ordinary (68% of gaps in the corpus are a single tick), so fading any sooner would have the ladder flickering" — the stated justification for `StaleAfterSeconds = 6.0` |
| `mudsharp.Tests/Fixtures/NpcHealthTrackingTests.cs:124` | "68% of miss-streaks in the corpus are exactly one" — degraded further, and pinning it against correction |

## Why it matters beyond the typo

At the true 42%, gaps of >= 3 ticks are `47/161` = **29%** — so nearly one health reading in three
trips the "stale" threshold, which is precisely the flicker the comment claims the threshold prevents.

**And it is the wrong statistic to begin with.** Those are *exchange* gaps (either side acting). The
health rung only updates on a **player landed blow**, which is a sparser distribution. So the constant
cannot be fixed by substituting 42% for 68% — it needs re-deriving from landed-blow gaps.

Note that `ParticipantRoster.cs` already uses the correct rate three lines below, for a different
constant: `UnknownAfterSeconds` is justified by `0.43^5 = 1.5%`, i.e. p=0.57 per tick. The same file
therefore carries two incompatible figures for one quantity, three lines apart.

## Scope

1. Correct the figure in all three places (safe, mechanical).
2. Re-derive `StaleAfterSeconds` from landed-blow gaps rather than exchange gaps. This is a judgement
about what the rail should show and wants deciding, not guessing.
3. `COMBAT-RAIL-SPEC.md` is marked LOCKED — amend it explicitly rather than silently.

## Provenance

Found by an adversarial audit of the `tools/combat/` corpus, 2026-08-23. Arithmetic independently
verified. This is the strongest single piece of evidence that agents reading that corpus are being
actively misled into shipping wrong constants.

Start with the histogram in tools/combat/MECHANICS-VERIFICATION.md and compare it with tools/combat/archive/TICK-PHASE-REVIEW.md, then inspect the locked wording in tools/combat/COMBAT-RAIL-SPEC.md and the comments in ParticipantRoster.cs and NpcHealthTrackingTests.cs. Correct the copied statistic in all three places and re-derive StaleAfterSeconds from landed-blow gaps, with the amended locked spec documenting the decision.

Written by the indexing model from the issue text.