Add mizer 3.3 announcement blog post

[gustavdelius]

Draft announcement for the mizer 3.3 release, built around four features that can be demonstrated rather than just described. Every number, message and figure in the post was produced by running the code against the current mizer source tree — nothing is invented output.

Structure

1. Is your model at its steady state? isSteady(), the new biomass drift line in summary() (0.014/year → 4.6/year after matchGrowth()), the message the match…() functions now emit, and plot(getSteadyResidual(params)).
2. Steady states you cannot reach by projecting. The North Sea model at effort 1.5, past its Hopf bifurcation: steadyNewton() finds the steady state, getStability() returns a complex leading eigenvalue 1.2359+0.847932i predicting a 7.4-year oscillation, and projectToSteady() settles onto a 5.4-year limit cycle. The gap between the linear and nonlinear period is explained rather than glossed over.
3. Scanning a model. Why measuring a quantity on an attractor requires knowing what the attractor is, then plotYieldVsF(NS_params, species = "Cod") with $F_{MSY}$ marked at 0.9, and the MizerScan summary.
4. Plots that know what they are showing. Array types, then the length-axis spectrum now carrying the resource and the total line that 3.2 silently dropped.

The final section covers the twelve guides and the skills they are generated from, and makes the point that matters for mizerAgents: setup_mizer_agent() now reads skills from the installed mizer, so upgrading mizer upgrades an agent's guidance with no mizerAgents release needed. The upgrading skill gets particular attention, since vignettes are not installed and that is exactly what an agent needs when a script breaks after an upgrade.

make-figures.R in the post directory regenerates all four figures.

Notes for review

• Date and version. Dated 2026-08-21, and the post says install.packages("mizer"). DESCRIPTION in the mizer repo is still at a development version, so both want settling before this goes live.
• Inline code links. filters/inline-code-links.lua resolves inline code against the installed mizer, which is still 3.2.x here, so the new names (scanModel(), isSteady(), getSteadyResidual(), plotYieldVsF()) are currently left unlinked. They will resolve once 3.3 is installed and the site is re-rendered.
• Only the new post's output is committed. A full quarto render on this machine also deleted docs/CNAME and the *_files/figure-html/ images of several older posts, whose regenerated HTML still references them — the same breakage that commit 7b83761 had to undo. All of that collateral was restored; this PR touches only the new post plus the site-level index.html, index.xml, listings.json, search.json and sitemap.xml, matching the file set of the 3.2 announcement PR (Add mizer 3.2 announcement blog post #5). Worth knowing that the render is not currently reproducible on a clean checkout.
• The accessor renaming (17 get-prefixed accessors superseded) is only a bullet in "Also in 3.3", since it does not demo well and the upgrading article covers it properly. Happy to give it more room.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BwrX4RLY5Fc7d6cu3fnJAQ

[gustavdelius]

Pushed a correction to the stability section (923e3c6).

Investigating the dt dependence I mentioned in the description turned out to matter for the post, not just for mizer. getStability() documents its continuous eigenvalues as independent of dt, but they are not, and the default dt = 1 is well outside the range where the discrete-to-continuous mapping has converged — filed as sizespectrum/mizer#556.

The post had been quoting the default-dt numbers, which overstate the growth rate by a factor of ~17 and give a predicted period of 7.4 years. Recomputed at dt = 0.1, where both branches of the analysis have converged:

• max_real_part 1.235902 → 0.07209386
• leading eigenvalue 1.235902+0.847932i → 0.07209386+1.216536i
• predicted period 7.4 → 5.16 years

That also changes the story for the better. The post explained the gap between the predicted 7.4 years and the observed 5.4-year limit cycle as nonlinearity reshaping the cycle, following vignette("dynamic_stability"). Most of that gap was discretisation error: the converged prediction of 5.2 years is close to the cycle the model actually settles on, so the linear analysis performs considerably better than the post claimed. The text now says so.

The code block passes dt = 0.1 explicitly, with a one-line parenthetical pointing at #556. If #556 is resolved by changing the default before 3.3 ships, that argument and the parenthetical should come out of the post. The vignette and the analyse-stability skill carry the same uncorrected numbers and the same nonlinearity explanation, so they need the same treatment.